Students are concerned: about 61% of the Class of 2026 students find current news about the job market somewhat or very pessimistic – up from 50% for the Class of 2024. High competition for jobs and lack of available jobs top the list of reasons for that pessimism, with the political climate and AI much more important factors this year relative to last.

Source: Handshake.

How should universities respond? There are obvious short-term actions such as helping students sharpen their job search materials, expand their job search, polish interview skills, etc. Universities can double-down on connecting with employers to identify additional opportunities for these newest graduates.

Looking beyond the Spring of 2026, universities should also be asking questions about how they can help students be even more career-ready – addressing what has come to be known as the ‘skills gap’.

In the fall of 2024, just after we launched Finding Equilibrium, we wrote a series of posts on the role universities, employers, and students can each play in better preparing graduates for the work world. Finding Equilibrium was new then, and many of our current readers weren’t yet with us. So, we thought we would revisit some of the key points we made in those posts given their continued relevance.

Before we do that, we will review where the job market for new college graduates stands currently.

How Does the Job Market for New College Graduates Look?

It HAS continued to soften – with unemployment for 22–27-year-old recent college graduates trending up since just after the pandemic to 4.8% in June. A recent and nuanced look at the unemployment data finds that while the unemployment rate for 22-27-year-old high school graduates remains well above those with a college degree, the gap between the two groups has narrowed dramatically to a level not seen since the late 1970s.

Source: Federal Reserve Bank of Cleveland.

This study also reports that the ‘job-finding rate’, the fraction of the unemployed who find a job on a monthly basis, has declined over time for college graduates and now is similar to that of high school students. It is now taking longer for young unemployed college-educated Americans to find jobs – about as long as a high school graduate. However, it is important to note that once a young college-educated person has a job, they have greater job security and enjoy substantial wage premiums relative to those without a degree. The issue is landing the job…

Source: Federal Reserve Bank of Cleveland.

So is AI taking these entry-level jobs? While different studies come to different conclusions, in aggregate, it does not look like AI is to blame. There is strong evidence the general weakness in new graduate employment has been in the works for more than two decades – well before the launch of generative AI. And, this recent downturn in hiring is taking place during a period of policy uncertainty (tariffs), a sharp drop in consumer confidence, and after a huge surge in hiring in some of the affected fields.

What About the Spring of 2026?

Looking forward, the National Association of Colleges and Employers (NACE) conducts semi-annual surveys of employer hiring intentions. The story told in the most recent study is aligned with the above discussion of the general employment situation. The November 2025 data show that 45% of the employers surveyed describe the job market as ‘fair’ – the worst characterization of the job market for new college graduates we have seen since 20-21.

Source: NACE.

That said, despite describing the job market as ‘fair’, 85.6% of the employers surveyed by NACE said they would be either increasing or maintaining hires. And, exploring the impact of AI on hiring for the class of 2026, NACE reported 13.3% of the employers said their jobs required AI skills and 10.5% of entry-level jobs include AI in job descriptions.

So, is the Class of 2026 facing some kind of nightmare when it comes to employment? The Fall 2025 NACE numbers don’t paint that dire a picture. And, The Economist calls current worries about the job market overstated, noting the “unemployment rate has been bouncing around 50-year lows”. In their view, while there are reasons for concern (unemployment is creeping up, some major employers are shedding jobs, consumer confidence is low), they don’t find reasons for a “jobs-pocalypse”: mainly because the US economy continues to roll along.

That said, some are more pessimistic and a recent Wall Street Journal article called the current job market ‘fragile’, pointing out firms have plans to grow without hiring. Hopefully the NACE/Economist take plays out. But, we won’t be surprised if the Spring 2026 NACE survey reports a steeper decline in hiring. Whether or not the ‘crisis’ label fits, it does look like many more graduates in the Class of 2026 will certainly see it that way…

Employers Care More About Experience than Grades

What are employers looking for from our students? The NACE study had some important insights: the use of GPA to screen job applicants declined (dramatically) from 73% in 2019 to 37% in 2023 but has ticked back up to 42% in 2025-2026. So, GPA is less important than it was pre-pandemic.

NACE also reports a list of factors considered by respondents beyond GPA when screening applicants. The list for those who do not screen by GPA is shown below (it looks similar for those employers who do screen by GPA). Anything that smacks of work experience stands out.

Source: NACE.

Taking this further, NACE asked employers, what is most important when choosing between equally qualified candidates. The student’s major matters, but everything else at the top of the list is focused on some kind of experience: internships, general work experience, leadership experience, co-curricular experience – GPA comes in at number 7. These results support arguments we made in earlier posts about the importance of accumulated experience to employers.

Source: NACE.

NACE doesn’t get into why the decline in screening for GPA or why GPA is not more important in selecting among candidates. A degree is enough? Post-pandemic grade inflation? Other factors just more important? (Looks like a topic for a future post…). Whatever the reason, experience and experiences matter – a lot. Universities need to be looking for every opportunity to help students build their experience portfolio as they progress through their academic program.

AI and the First Job

AI is having a major impact on the hiring process. Candidates use AI to generate application materials making traditional cover letters (which now all sound alike) of little value in screening candidates. Firms are responding by using AI to screen applications for specific skills, giving tests and evaluations where the candidate can’t use AI, and focusing more on specific target schools and/or personal recommendations rather than standard application packets. Students need to know that AI may be making the job application process easier for them, but may well be making the likelihood of landing a job harder.

While AI may not yet be taking jobs, it is starting to impact what graduates do in that first job. Many of the tasks that a new hire performs are somewhat basic, requiring codified, explicit knowledge such as information gathering, basic data analysis, report writing, presentation development, etc. – the kind of stuff AI is already pretty good at. So, students will be expected to know how to use AI to perform those tasks.

Using AI to perform the basic tasks new hires once performed creates a different problem: those basic tasks are a form of training for the new hire, learning the industry, the customers, the firm, understanding how more experienced workers approach problems. All of this builds the tacit knowledge that firms really want! As reported in NACE, employers are looking for experience in who they hire. And, once hired, new employees need to be very deliberate in seeking out and developing such tacit knowledge in that first job.

Compounding this experience issue is the current move to eliminate middle management positions in organizations, increasing the number of direct reports any manager has. One study reported business had 1 manager for every 5 employees in 2017, increasing to one for every 15 employees in 2023. More direct reports mean managers have less time for coaching, mentoring, building personal relationships with direct reports and so on. One implication: a workplace where “only high-performing self-starters thrive”.

Universities need to recognize that the work world their graduates are stepping into is changing and doing as much as they can to help them prepare for those changes.

Addressing The Skills Gap

Looking beyond the Class of 2026, what can universities do to better prepare their graduates for the work world? We wrote extensively on this question in the fall of 2024. If you were with us then, the rest of this post is a summary of our ideas. If not (or you want a review), read on!

We took a careful look at the disconnect between what employers want and what new graduates bring to the work world. Surveys consistently show that new graduates are falling short of employer expectations on professional skills such as leadership, communications, adaptability, problem-solving, and so on. Likewise, students in many cases don’t bring the mindset or disposition employers want to their first job – such attributes as drive/work ethic, motivation and initiative, and resilience and persistence. (Disciplinary knowledge gaps aren’t as well documented, perhaps because they are not as carefully explored.)

Who is to blame for the skills gap? Universities? Employers? Students? We don’t see such finger pointing as helpful – stepping back, we believe four important factors are driving the skills gap: information (not enough of the right information available at the right time); incentives (needed to produce career-ready graduates aren’t there); structures (organizational structures don’t support needed accountability and responsibility); and generational differences (the characteristics of Gen Z affect career preparation).

We’ll look at each of these areas in turn.

Better Information

Employers don’t know enough about curricula and what universities are doing to help students build career skills. Faculty don’t know enough about what the work world expects from their students. Students don’t know enough about available careers, what it takes to be successful in those careers, and what choices they need to make to best prepare themselves for the work world.

We offered a four-point plan to address these information issues.

• Collect and publicly report placement data for all students in order that students, families, and universities are much more informed about career outcomes across academic programs.

• Capture co-curricular participation on transcripts - many of the professional skills in demand by employers are developed through engagement in co-curricular experiences. Capturing these experiences on transcripts signals to students they are important and to employers that a student has engaged in them.

• Partner with state agencies to provide data on career progression and earnings throughout the professional life of a graduate. Starting salaries are just that – a start. Students, families, and universities need to understand how college impacts student financial outcomes over the course of the graduate’s career.

• Develop forward-looking employment forecasts - more employer engagement is needed to help universities understand where the job market is going. Universities need to be preparing students for the jobs to come, not for the jobs that were.

Given rapidly advancing AI capabilities, addressing these information issues is more important than ever…

Incentives that Support Career Readiness

Are incentives in place to ensure faculty are teaching in a way that helps prepare students for their professional lives and to ensure curricula remain relevant and aligned with the work world? In general, we don’t believe either of these questions can be answered in the affirmative. We offered a set of ideas including 1) elevating the importance of teaching on campus; 2) preparing and supporting faculty to improve their teaching; 3) increasing rewards for innovation and impact in instruction, and 4) getting serious about curricula review and design.

What about students? How do we improve their incentive to do the work and focus on the courses and experiences that will help prepare them for a career – trading off less fun today for something that may benefit their career benefit years in the future? Most important here: relentlessly connecting the choices students make today with the benefits they will receive a few years hence.

Employers want students who have been educated to meet their specific job requirements – but training for a specific employer’s specific job is not what a university education is all about. So, universities should listen to employers – but not too much. Making it easier for employers to inform decisions about curricula, call out the skills gaps they see, and identify initiatives that will help students be more career ready is important.

Such engagement is powerful as long as the university keeps the student and their career path at the center of their decisions. Some ways to do this: work with industry or trade associations to better understand industry (as opposed to a specific employers) talent needs; bring employers into micro-internships and project-based courses; pair such employer engagements with an offer of skills-certifications for students; and studying all these student-employer engagements to draw campus-wide insights on changing workforce demands.

Supporting Gen Z

Some blame the skills gap on Gen Z – a generation they see lacking motivation, work ethic, resilience and grit, focus… Maybe some of this is true, but so are a host of positive traits: Gen Z has been characterized caring deeply about others, being highly collaborative, valuing flexibility and relevance… The point: universities must meet Gen Z where they are, not cast aspersions about who they are not.

Ideas here include advising focused on assessing each student’s work and leadership experiences and individualized counseling to steer students into curricular and co-curricular options that best complement strengths and address preparation gaps. Making sure faculty understand generational differences is important in modifying pedagogy to fit Gen Z learning styles. Universities should help employers understand this new generation – helping them see how Gen Z is different from (and similar to) previous generations.

In the final post in the skills gap series, we addressed a fundamental issue on university campuses: who specifically is responsible for the employability of a student? Every university offers a myriad of resources to support student preparation for the work world – career offices, advisors of all forms, mentoring programs, career fairs, professional development certificates, programs, workshops, and seminars, among many others.

Yet, it falls on the student to find their way through this maze of support to determine which of these many, many options make sense for them and how they can best use them. In essence, students are their own general contractors when it comes to building a career-ready education – and some students are far better equipped to play this role than others.

Getting better here involves actions such as mapping student contact points with available resources, simplifying and streamlining support resources where possible, sharing relevant information across student support offices, reorganizing resources and offices where needed – and importantly, looking for ways to integrate employability into the curricula.

Some Final Thoughts

A soft job market for graduates is nothing new for universities, but it certainly will be new for members of the Class of 2026. So, these students deserve our very best in helping them secure that first job. That said, every university should also be revisiting the way they think about and prepare their students for a successful career. It’s not only the right thing to do, in an environment where tuition revenue will become increasingly important, delivering career-ready graduates may well be the ticket to keeping the doors open…

What’s Next?

After a hiatus over the holiday (and giving you a break!), we’ll be back in mid-January. Some of the topics on our minds for Spring 2026 include the enrollment cliff, college athletics, low-income students, non-degree credentials, and fostering teaching excellence. That should keep us busy, but we have little doubt something else will pop up that we decide to pop off about! Maybe we will even get our podcast going in the Spring – the intentions are there…

We sure wish you, your families, and your friends the very best of the coming holidays. Thank you for reading Finding Equilibrium and we look forward to exchanging ideas with you in 2026!

Subscribe now

Research assistance provided by Marley Heritier.

“Finding Equilibrium” is coauthored by Jay Akridge, Professor of Agricultural Economics, Trustee Chair in Teaching and Learning Excellence, and Provost Emeritus at Purdue University and David Hummels, Distinguished Professor of Economics and Dean Emeritus at the Daniels School of Business at Purdue.

When we started talking about this issue last summer, we didn’t know what we thought universities should do. So we wrote the previous four essays in this series to figure it out, starting from these questions: what can we learn from historical analogues, what is the current evidence on AI adoption, and how are AI tools likely to remain distinct from human cognition?

What Have We Learned from Our First Four Posts?

1. Generative AI tools can predict but they don’t understand causal mechanisms.

The key to solving deeper problems lies in answering “why” something works, as opposed to “how” something works. Because generative AI cannot form its own “world models” that describe causal mechanisms it cannot effectively generate its own “why’s” or evaluate whether its “how’s” are consistent with the “why’s” we already understand. Deep disciplinary knowledge and strong frameworks for understanding the world are required to extract value from AI.

2. Generative AI produces output subject to serious quality verification problems.

When producing statistically likely strings of tokens, generative AI is drawing on an entire internet of text, which includes both authoritative and less reliable sources, as well as some teenager’s Reddit thread. It is also given to producing fluent but ultimately meaningless text (“workslop”) and to hallucinating statistically likely, but untrue, strings of tokens. Workers with an ability to rapidly discern source validity and distinguish accurate from merely plausible-sounding output will be particularly valuable to firms.

3. Firms’ adoption of AI tools to-date is widespread but largely experimental.

Firms don’t yet know how to use these tools and time- and cost-savings to date seem modest, which is consistent with earlier experiences with transformative technologies. If our goal is to make students ready for an AI-augmented work world, this uncertainty makes it hard to know precisely what universities should be teaching students about AI-use.

4. Two clear patterns for AI-use *are* emerging: information retrieval and tasks that require explicit, codified knowledge.

Many entry-level jobs intensively involve information retrieval: gathering information about goods, services, sources; responding to customer problems or inquiries. Others consist of essentially algorithmic tasks, following a predetermined process for generating an outcome of value. The challenge for universities is that too many of the assignments we give students are precisely the sorts of things generative AI is already good at, which means that many of the skills developed by our newest graduates are automatable by these tools.

5. Most jobs and workflows consist of a large set of tasks.

Even in cases where AI tools might be useful for replacing some specific tasks, firms have to figure out methods for integrating these replaceable tasks into jobs and a larger workflow. This will slow adoption but also put a premium on workers if they can use their AI-enhanced ability to rapidly and accurately solve a single well-defined problem and then integrate that solution into a more complicated problem.

6. The real value of new technologies lies not in reducing labor costs, but in creating new opportunities.

The most transformative new technologies are utilized by creative entrepreneurs/firms to generate entirely new classes of products and services that consumers didn’t even know they wanted. This turns the employer problem from “which workers can I layoff?” to “what new value can these workers create with the help of generative AI tools?” Universities need to focus more on preparing students to play a role in answering that second question.

Where are Students on AI?

Before we lay out our recommended actions, one more important point: universities must understand student attitudes about - and use of AI - as they put together their AI strategy.

Many students are nervous about AI and their future employment. For the class of 2026, about 61% find current news about the job market makes them either somewhat or very pessimistic about their career prospects – up from 50% for the Class of 2024. The political climate and generative AI were the two fastest growing reasons for that pessimism.

AI use by Gen Z is both widespread and frequent – in a recent survey 79% said they use generative AI products, with 47% saying they use these tools on a regular basis. But just because they use generative AI does not mean they are comfortable with the technology. Perhaps because it is so new and because students don’t yet see clear rules/guidance from schools and employers on how best to use the tools, they have real reservations.

These reservations are shown in the two charts below: while Gen Z sees AI as helping them find information easier, complete work faster, and learn faster, they are also very concerned about AI’s impact on their cognitive abilities in searching for and using information as well as coming up with new ideas.

Source: Voices of Gen Z

Source: Voices of Gen Z

Some Curricular Implications for Universities

Given where we are (and where our students are) with AI during this time of transition, what steps can universities take to best prepare their students for the evolving work world? Certainly one element of a university’s response will be adapting learning assessments to ensure that they are more AI-proof (more below on this). But that is only a start. Our thinking is well-captured in this Washington Post Op-Ed:

“Universities have been successfully delivering fresh talent to companies for decades. Now, in the face of AI, they need to reimagine what they do. This doesn’t mean fighting chatbot-enabled cheating. The bigger threat to the higher education system as we know it is that many of the skills colleges develop, such as the ability to do basic research or write coherent prose, will be devalued in the job market as AI takes over. On the other hand, skills that universities cultivate – such as critical thinking and analysis of text – could empower graduates to use and evaluate AI tools.”

More concretely, what does this look like?

1. More Experiential Learning

In the before times, firms used extended employment training periods where new graduates engaged in relatively low-valued information processing jobs. Then in the course of performing assigned tasks, these young employees developed the capabilities that prepared them to take on more sophisticated and integrative roles in the firm.

If our recent graduates produce work that is too similar to what generative AI can produce today our question becomes: how do we make a 22-year-old new college graduate look like an experienced 27-year-old in the eyes of employers?

Central to the answer is a much bigger focus on intensive experiential learning opportunities. Too many of our courses consist of teaching students to provide algorithmic solutions to a set of narrowly defined problems that instructors pose. (Here is a test. If students are asked a question that has a single “right” answer, then it follows that this answer must be arrived at by memorizing a fact or applying some algorithm to a set of initial conditions.) These capabilities and assessments are still necessary – it is difficult to learn conceptual frameworks without practicing their use – but they cannot be the endpoint.

Students need more repeated, increasingly sophisticated, opportunities to (a) define problems; (b) hypothesize potential solutions; (c) evaluate and choose from among that set of solutions; (d) communicate their solutions to peers and to faculty; (e) interpret and evaluate feedback to improve their solution. Each piece of this process requires understanding the disciplinary solution frameworks we teach now, but also understanding which to use, how to work with a team of peers, and how to integrate multiple sources of information and analysis into an integrated “product” – for a messy, open-ended, ill-defined problem.

2. Generating New Knowledge/New Value

As Pablo Picasso once said: “Computers are useless. They only give you answers.” We spend far too little time getting students to ask their own questions, arrive at their own solutions, and defend what value these solutions have in different contexts. The typical excuse for that has been that it takes too long to get students close to the knowledge frontier; asking them to push it forward is a job for a PhD student.

Maybe. But perhaps thoughtful use of genAI in support of experiential learning might make it possible to accelerate the process of moving to the frontier on some dimension. We will be experimenting this spring with using genAI to facilitate independent studies with extremely aggressive timelines and deliverables. We’ll let you know how that goes.

3. Evaluating Source Validity

The advent of the internet meant that students could bypass the usual gatekeepers (book and journal editors, responsible media outlets, librarians,…) of information and rely on an increasingly diverse set of self-chosen sources when collecting and processing information. The field of “Information Literacy” arose in part to provide students needed tools to discern valid from invalid sources. (Oh, how quaint the handwringing over Wikipedia now seems!)

Generative AI puts this problem on steroids for three reasons. One, it is using some inscrutable predictive process for generating text from its training data in a way that obscures the underlying source of “facts” or “analysis”. Second, when prompted to provide sources, genAI is likely to report a mix of authoritative and non-authoritative sources. Third, in many cases there is no “source” in the sense we usually mean (there is a mapping between a particular claim and a specific place where that claim was made). That includes both statistical amalgamation of sources and the possibility that there is no place whatsoever where this specific claim was ever made.

This calls, we think, for a major re-emphasis on information literacy for the AI-age. Essentially teaching students not only to distinguish authoritative sources, but to become source detectives: where is this claim actually coming from and should I trust it?

4. Epistemology: What is Evidence?

A deeper question that lies behind source validity is the question of how we know the things we know. Each academic discipline has a set of questions it tries to answer, and approaches to finding these answers. The discipline has within it particular standards of evidence that we use to evaluate claims.

In our view it is necessary to get much more explicit in teaching students about these approaches. How we define questions, how we arrive at answers, how we discern whether an answer meets a standard of evidence. Faculty may think they are already doing that. Our guess is that if you actually asked students what they understand about this process, most would be lost at sea in trying to answer.

Being more explicit would help to demystify the educational process and help motivate students in their learning. It also helps students to apply these same approaches when engaging in independent and experiential learning. When they leave the university, students often complain that their jobs don’t use what they learned in school. We maintain that they don’t see the connections because we do a poor job of helping them to see the connections.

5. Ensure Students Know the Path from the Start

We can already hear the howls from students – just let us use the AI tools already! Jeez…what’s with all this stuff about disciplines, frameworks, questions, sources…we know how to prompt AI to get what we need! Well, with genAI, that is sort of like getting into the cockpit for your first solo flight without ever completing any of the on-the-ground training and instructor-assisted flights…

In the spirit of motivating learning, universities will need to help students make explicit connections from day 1 about why building a world view, understanding a discipline, and being information literate are essential if the power of genAI is to be realized. And why experiential learning and the power that flows from asking and answering their own questions will make them much more valuable when they enter the work world.

What are the Challenges We Face in Pursuing this Approach?

Reflecting on these five points, isn’t this stuff that universities should be doing anyway? Well, yeah…go back far enough, and most universities probably were. But cost pressures, student numbers, student preparation, and a focus on other priorities have taken us down a different path – and now many of the ways we teach and assess do not develop the kind of capabilities demanded in an AI-enabled world.

In response, some universities are launching surface level AI initiatives, an AI-course or two and administrative edicts to put AI to work. We believe it is more important to double down on the human skills that AI cannot replicate. Ensuring that the type of learning environment needed is campus-wide and not just tucked into an honors program/honors courses is a major lift. What are some of the issues that will need to be addressed (and what can be done about them)?

1. Creating AI-proof Assessments

A variety of ideas are being put forward to ensure that students do not outsource required work to AI. One challenge here is these approaches are likely to be extremely specific to the individual course and traditional style of assessment, and this is likely to evolve rapidly. Last Fall economics professors could get around “text-only” chatbot responses by asking students to use graphs to demonstrate concept mastery. But in the span of a few months, these tools have gotten quite good at drawing and analyzing graphs. This points to a near future where almost any outside-the-classroom assignment can be significantly AI-assisted. The deeper question we have to answer is: what experiences and what assessments will best build learning and capabilities under the assumption that students and their future-selves will always have these tools at hand?

2. Experiential Learning is Expensive

Putting 500 students in a big lecture hall with three-computer graded exams and some TA graded homework is a good way to drive instructional cost per student down. Finding creative ways to deliver the kinds of experiential learning we describe above cost effectively will take some work. Smaller class sizes, more time on assessment, resources for faculty (and student) support… Maybe AI can be used to ‘teach’ some topics where building codified, explicit knowledge is the objective, freeing up resources to invest in courses building deeper learning skills…maybe… Active learning approaches are a big step in this direction, moving lecture material out of the classroom and putting classroom time to work on problem sets, case studies, team discussions/debates, etc. Course-based undergraduate research experiences (CUREs) where students are actively engaged in research experiences inside a course with a focus on engaging in inquiry with unknown outcomes is another approach which cultivates the kind of thought-processes we describe above – and has the potential to be scaled.

3. Do Students Want These Experiences?

The kind of critical thinking, data literate, disciplinary education we are describing will be real work for most students. They are accustomed to regurgitating on exams and homeworks whatever the professor has been professing; engaging in straightforward information gathering exercises; and performing carefully structured activities with equally carefully structured deliverables. The question then becomes: are students really ready for this type of open-ended, inquiry-based instruction? Some students are and would thrive – as many already do in Honors curricula. But, and there is a big *but*, helping the broader student body understand this is where we are going and then making sure they have the support to be successful is going to take careful thought.

4. Centralization of AI Policies/Approaches

Students are looking for consistent, university-wide guidance as to acceptable use so they can put AI to work in creative ways. We as coauthors disagree on whether or not this is possible or needed (maybe a Provost vs. Dean thing…):

• Jay: Universities will need – and can develop – some general AI guidance for a campus that preserves local latitude (a version of this is playing out now as the federal government and states duke it out over who will determine AI policy). Clarifying issues about what is expected (every course has an AI policy,…), fair use (students must follow class policy and in absence of policy, acknowledge how AI is used,…) will be important to at least define the campus AI landscape and manage some chaos/mischief. And, every student should get some exposure to a basic set of AI applications/skills (research, brainstorming, written communications, come to mind) and some of the ethical/legal uses of AI (however they are defined at that moment in time) – maybe during first-year orientation.

• David: I’m in that states’ rights camp! I agree that every student needs exposure to AI applications/skills, but I think these are most useful in the context of discipline-specific engagement with those tools. Clear course-level policies are important, but course-level (as opposed to campus level) policies imply situations in which an AI-practice that is praised in one course could be treated as academic dishonesty in another. I am extremely skeptical that universities can create a centralized policy that provides meaningful common guidance and accounts for local latitude, especially given the unprecedented speed with which this technology is evolving. As long as the course-by-course rules are clear, students can adapt, in the same way they have to adapt their behavior to context in every other aspect of life. Just read the damn syllabus!

5. Supporting Faculty, but How?

Given genAI is still relatively early in the adoption cycle, how exactly do we help faculty move in the direction we describe above? Providing insights on more AI-proof assignments is a starting point – but that won’t necessarily get at developing the mindsets and skills needed. It is time for experiments and pilots, extracting lessons from those experiments and pilots, and then providing the resources for campus-wide rollouts of approaches that work. This is what firms are doing with AI now – running lots of experiments to figure out what works – why not universities? Faculty will deliver the kinds of innovation needed, but universities can get far more systematic about scaling these innovations through ‘stage-gate’ type approaches to managing innovation used in the commercial world. Using a stage-gate model in a university setting, innovations that work at the course level are flagged and explored further to see how they might work at a Department, then College, then campus level, as opposed to just sending the faculty member a note of congratulations for making their course better.

6. Staying Current

We highlighted in an earlier series of posts the importance of engaging with employers to better understand the work world. Employers are aggressively implementing AI and exploring where it pays and where it doesn’t. Universities (and their graduates) can’t afford to wait for some steady state – they need to know what is going on where, how what employers want from entry level hires is changing, and what all this means for the work environment new college graduates are moving into. Staying current with AI deployment is not a one and done deal: it is going to take on-going engagement with employers, likely at the college/department level, to stay anywhere close to what is going on in the work world.

7. Cost of and Access to AI Tools

We’re all sort of taking it for granted that these AI tools will continue to be available for free student use in the future - and that may not be right. The primary developers are absolutely hemorrhaging money. The “do a glorified Google search and aggregate the results” types of queries are very low cost to provide. The “thinking” queries, especially for multiple step functions start to get really expensive… orders of magnitude more expensive. Right now those are heavily subsidized to build a user base but we would expect increasing efforts to monetize and then a shakeout with the remaining players using their newfound monopoly power to price. What that means for students we don’t know… but we could imagine a very uneven future where some students are using the low-rent chatbots and others have the gold-plated premium bots. Unless the university springs for the additional expense…

8. AI and Student Mental Health

We also didn’t talk at all about AI psychoses. But some disconcertingly high fraction of openAI users show evidence of psychoses, and suicidal ideation. An important question, once universities start actively encouraging AI use, is what happens when some vulnerable students go off the deep end with tools we provided/encouraged? Student mental health issues associated with AI need to be on a university’s radar.

A Final Thought as We Close This Series on AI

As we wrap up our series on AI, there is still tremendous uncertainty as to how all of this will play out. Universities can fully engage in the AI adoption process and resulting transitions – including helping firms put AI to work productively. Or, they can watch from the sidelines as their faculty, staff, and students try to figure it all out in real time. We would suggest the former – the latter looks to be one more step toward being declared irrelevant.

Next Week

We’ll end our year (and send you into the holiday break) with some thoughts on the employment situation for the Class of 2026 and how universities can better support their students – both in the coming soft job market and longer term in helping them make the step into the work world as short as possible. As always, thanks for reading Finding Equilibrium!

Subscribe now

Research assistance provided by Marley Heritier.

“Finding Equilibrium” is coauthored by Jay Akridge, Professor of Agricultural Economics, Trustee Chair in Teaching and Learning Excellence, and Provost Emeritus at Purdue University and David Hummels, Distinguished Professor of Economics and Dean Emeritus at the Daniels School of Business at Purdue.

Let’s start with some basics about how humans reason (drawing heavily from Erik Larson’s book The Myth of Artificial Intelligence).

Human Reasoning and Inference

Reasoning allows humans to interpret information, solve problems, and make decisions. Philosophers and cognitive scientists describe three primary types of reasoning: deduction, induction, and abduction. A central thesis of Larson’s book is that, while abduction is a critical tool in the human arsenal of reasoning, it lays fundamentally beyond the reach of generative AI. This greatly limits what AI is capable of and creates an enduring space for human ingenuity.

Perhaps many readers will be familiar with deduction and induction, but not abduction. Below I lay out how abduction connects to a series of related concepts, including “tacit knowledge”, “heuristics”, “crystallized intelligence” and the whole “10,000 hours makes an expert” idea/myth. But first, let me briefly describe deduction and induction to provide a contrast.

Deductive Reasoning

Deduction involves deriving specific conclusions from general premises. The strength of deduction is that valid premises lead to necessarily true conclusions and that reasoning is transparent and verifiable. The weakness lies in the premises and the narrowness of application. Premises must be true and meaningful to tell us something important about the phenomenon of interest. Deduction is particularly limited when outcomes are multicausal. Which is another way of saying that deductions might be able to work in one causal direction but not the other.

Consider this example: If it rains, the ground is wet. It rained. Therefore, the ground is wet.

While this statement is a valid deduction, it is not deductively valid to go in the other direction and say: the ground is wet, therefore it rained. Why? There could be many reasons the ground is wet. Ground wetness is multi-causal.

Economics uses mathematical models as a form of deductive reasoning. We lay out a set of premises (assumptions about consumer preferences and firms’ technology), and a solution algorithm for combining those assumptions (e.g. market equilibrium) and then derive conclusions that necessarily follow from the mathematics of the model. Within the confines of the thought experiment, this can provide powerful insights, including predictions of things that have not yet happened.

Indeed, many academic disciplines develop frameworks for analysis that have a similar deductive reasoning structure. General premises lead to specific conclusions that can be used to explore possibilities that are beyond our historical experience. For example: what happens when you create a chain reaction in a pile of fissile radioactive material, or when you launch a rocket into the vacuum of space?

Inductive Reasoning

Induction involves observing specific instances and then inferring general principles. Unlike deduction, induction results in probabilistic beliefs – predictions that are likely, but not certain.

Example: Every swan you’ve seen is white, and from this, you induce, “All swans are white.”

This inference seems reasonable, but there are, in fact, black swans. They are just statistically unlikely. The strength of induction lies in its ability to discover patterns and assess hypotheses given data. The weakness lies in what data are actually available. The field of statistics focuses on what we can reasonably learn from a given data sample, and provides a long list of the perils of inference that stem from those limitations. Large Language Models exclusively use induction.

Abductive Reasoning

Abduction, sometimes called “inference to the best explanation”, helps us generate hypotheses to explain observed phenomena when data are incomplete, ambiguous, or inconsistent. Unlike deduction and induction, abduction involves proposing plausible explanations based on prior knowledge, context, and intuition rather than strict logical inference. Or to be less generous: guessing.

Example: you wake up and see that the ground is wet. You hypothesize, “It rained overnight,”

This inference is plausible, but there could be other explanations given the multi-causal nature of ground wetness. Perhaps you left your sprinklers running overnight. But rather than working through all these possibilities, you select a hypothesis that is the most consistent with your background knowledge and observations.

(To see how this differs from an LLM, ask a friend and then your favorite chatbot, “The ground is wet, what do you infer from that?”)

The strength of abduction is that it enables rapid hypothesis formation in complex or uncertain conditions. It facilitates discovery, and diagnosis, and supports strategic decision-making when evidence is incomplete or ambiguous.

Rapid matters. We do not need a lengthy pause to assess all possible causes; we form immediate best guesses and have the ability to quickly amend those guesses. You may believe initially it rained, but you could very quickly modify this inference by noticing that, in fact, the sprinklers are still running.

The weakness of abduction is that the “best” explanation is always probabilistic, which can mean even best guesses can be wrong. Psychologists and behavioral economists have documented a long list of cognitive biases in our reasoning, particularly when we make rapid judgements.

Larson argues that abduction is a human superpower that generative AI lacks. If that’s right, one way to accentuate our cognitive differences relative to LLMs is to get even better at abduction.

Experience… and 10,000 Hours

Humans have two ways to improve their skills at abduction, that is, to get better at formulating hypotheses about what has happened or is likely to happen, and to use that to drive decision making.

One approach is to simply experience the world. We see that in the example of small children who very quickly learn about cause and effect, about physical properties of liquids and solids, about space and time. It is also true as we experience the world in specific contexts that require and develop highly specialized knowledge.

In Outliers, Malcolm Gladwell popularized an idea that 10,000 hours of practice are needed to develop expertise in particular fields. We shouldn’t take this notion too seriously, especially the idea that you only need 10,000 hours as opposed to many other factors that contribute to expertise. But it does capture a fundamental truth: people learn by doing. And if they are attentive and diligent, they can learn a lot.

Consider two personal examples. When I started doing research, I felt lost. I didn’t know how to identify an interesting question. To search the literature. To write down an economic model. To find data and test that model’s predictions. To write it up in a professional style. To communicate findings effectively in a research presentation. Or really, where to start on any of it.

Initial efforts at research were slow, full of wrong steps, and the output was not very good. Over time I progressed to the point where all of these steps happen more quickly and intuitively. Mistakes still happen, and things *still* take too long, but wrong steps are much less likely. I’ve also developed a kind of intuition for what will happen (in models, in data) before I work out the details. My younger self would be astonished at how often that intuition is right.

About 11 years ago I was asked to step in as interim Dean at Purdue’s business school. I agreed and was quickly overwhelmed. Not just by the scope of the job, and the number of things I needed to do, but by the scope of the things I needed to know before I could do anything. My decision making was paralyzed by rumination – did I know enough about what effect, and what unintended consequences, would result from any action I took? I rarely made decisions on the spot because I needed time to think, at least overnight, and sometimes longer. Really tough decisions could drag on for weeks or months. Indecision does not a good leader make!

But then something started to happen. I started to know what I needed to know. I started to understand the systems I was working in and the people I was working with. And my decision cycle sped up dramatically. I could be asked a question at the beginning of a 30-minute meeting and be comfortable making a decision by the end of it. I think the quality of my quick decisions arrived at later in my Deanship were better than the slow and ruminative decisions I made earlier.

What was happening? Through experience I was building world models of the behavior of the people and the system around me. I was using those models to form better abductions.

Now, here’s the challenge. If you ask me to describe exactly how I intuit the right way to progress through a research paper, or the right way to collect and analyze information or weigh consequences to make leadership decisions, I would struggle to do that. Much of what I understand about these things is what we call “tacit knowledge”.

Tacit Knowledge and Crystallized Intelligence

Polanyi’s Paradox, “We know more than we can tell,” nicely captures how much of our expertise is implicit and cannot simply be codified. Experts rely heavily on tacit knowledge stored in mental models of how the world works. These internally-held models allow for rapid, context-rich hypothesis formation and action. They might be impossible to articulate explicitly but are central to effective reasoning.

Now let’s call back to LLMs: they don’t “know” things, they simply predict tokens, words, based on statistically likely patterns they have already seen in their training data. While this training data is vast, it consists of things that are written down. If Polyani is right that we know more than we can tell, doesn’t it follow then that most of what we know as humans is invisible to LLMs? Which means that requests for LLMs to provide context-dependent judgement and decision-making will result in them providing responses that are outside the scope of their training data.

(Notably, recent efforts to train generative AI have started to move away from language based training sets to video based training in order to induce a deeper understanding of the physical world. But this is quite distinct from understanding how experience leads to judgement.)

This series of insights also links back to an interesting discussion I found a few years ago in Arthur Brooks’ book From Strength to Strength aimed at people who had moved on from positions of great responsibility and who were seeking meaningful work for a last professional chapter. Ahem.

Brooks describes the difference between fluid intelligence and crystallized intelligence. Fluid intelligence involves basic processes of reasoning that depend minimally on prior learning but that flow into many cognitive tasks. Crystallized intelligence is the “..application of learned procedures and knowledge, and depends heavily on experience and acculturation.” It “precipitates out of experience,” Brooks uses this as a jumping off point to argue (effectively!) for a richer kind of purposeful work to be found at later ages.

Or to paraphrase. Brooks argues that highly experienced individuals build tacit knowledge that helps them to create extremely rich world models of physical and/or human systems to engage in highly accurate abduction. A key to all that is remembering and applying lessons from experience.

Memories…

Memory shapes who we are and how we think in powerful ways. This was brought home to me 11 years ago when I crashed my bicycle and suffered a traumatic brain injury. For the better part of a day, I was not able to form new memories, which led to this disconcerting situation.

I would wake from sleep and say “my head hurts, where am I?” and my hospital room visitors would tell me. I would ask what happened, then ask if my riding partner Kip had cut me off. (He swears he didn’t but… I know Kip, I know.) I went through a sequence of 6 questions, then fell back asleep. Only to wake a few minutes later and repeat the exact same questions in the exact same words in the exact same sequence. Over and over again all day long. Funny at first. And then, not so funny.

Thankfully my brain function began to slowly recover the next day and shortly thereafter I agreed to serve as interim Dean. Make of that what you will.

But for that one day, I was exactly like a large language model. I had a pre-programmed set of parameters in my brain that mechanistically led me to generate the same questions over and over. I couldn’t progress because memories wouldn’t form in my damaged brain.

LLMs offer an amazing trick. They encode the internet’s knowledge in an unthinkably complex set of weights, and can use that to answer questions, produce essays, summarize knowledge. But those weights don’t change, which means they don’t remember and so they cannot progress.

Programmers have tried to work around this limitation by creating large “context” windows as a kind of short-term memory (if you’ve got an open chat going with ChatGPT or Gemini with a lot of back and forth…that’s the context window). But these are not the same thing as learning using memory. Once the context window is closed, the model reverts back to its base state. Just like what happened to me every time I feel asleep in that hospital bed.

Even while the context window is open, the usefulness of an LLM’s short-term memory is limited by three factors.

• LLMs lose track of information in the middle of a long context window.

• The amount of working memory is a vanishingly small fraction of the larger “all the internet” training data they originally learned from.

• The processing time slows, and the amount of memory and processing power needed to use a context window grows in proportion to the square of the context window’s size. A ten-fold increase in context requires a 100-fold increase in compute.

A three-year old uses memory to learn and adapt. Experienced executives use their tacit knowledge stored in memory to make subtle context-specific judgements. LLMs can do neither.

The Role of Formal Learning in Academic Disciplines

A second approach to improving your abduction, your skills at formulating hypotheses about what has happened or is likely to happen, is to study past human knowledge. Various academic disciplines represent different approaches to formally understanding the world, including structured causal models of how things work in different domains.

Some of this understanding (physics, chemistry, biology) connects to the physical world. Some (engineering) connects to how to manipulate that physical world. Some of it (economics, psychology, sociology, anthropology) to human behavior. Some of it (education, law, medicine, business) to human systems or professional practice.

The rise of generative AI does not mean this collected understanding is less important now than before. In fact, genAI makes these disciplines more important. Let’s explore why.

Causal Explanations

All academic areas of study encode what Marcus would call world models, frameworks for understanding and manipulating the physical or human world. A key element of the world models emerging from academic disciplines is a causal understanding of not just how things work, but why things work.

And the why – the causal mechanism – is really what matters. Think back to the study I described in the last essay that showed LLM-type models could accurately describe how planets moved but not why. Without understanding the why, predictive models are subject to serious out-of-sample failures. That is particularly relevant when we seek to understand physical or human behavior we haven’t already seen and that may differ depending on context.

In contrast, academic disciplines seek to build analytical frameworks that abstract from the specific details of a particular situation in order to identify broader principles that apply to many situations. That includes situations whose contexts are different from what we have previously seen.

Newton’s laws of motion and gravitation, Maxwell’s descriptions of electromagnetism, the laws of thermodynamics, conservation of mass, evolution by natural selection… all these very general frameworks apply (almost) universally.

In my field of economics, we might think about the role of scarcity, of opportunity costs, of incentives, of comparing marginal costs and marginal benefits. These are analytical frameworks that give one insight across an enormous number of situations because they get at casual mechanisms. When economists try to understand something we haven’t seen before, we reach first for these tools to understand them.

These are all examples of using known frameworks to understand specific situations we have not yet encountered. Understanding causal mechanisms is also critical for extending and deepening our understanding beyond frameworks we currently know. In every academic discipline, discovery builds upon discovery, making use of insight to generate new insight. To do this you need to know why something works to extend the underlying principle.

The same is true of the use of scientific knowledge in generating technological knowledge. Understanding Newton’s second law precedes the design of jet engines (expelling hot compressed gases backwards creates forward thrust that propels an aircraft). Understanding the Ideal Gas Laws precedes understanding why propellers don’t work in the very thin air at high altitudes and why you’d need a jet engine instead.

As far as we can tell, genAI lacks the capability to induce its own understanding of any of these, or the underlying casual mechanisms, which make it incapable of advancing the state of knowledge on its own.

Identifying and Preventing Predictive Bias

Predictive algorithms, like LLMs, are subject to a variety of predictive biases. This has gotten attention in the context of things like racial bias in sentencing, or medical diagnoses, but it’s a far more general principle. The issue is that the most accurate predictor (i.e. the one that makes the smallest errors on average) can also be statistically biased toward a particular outcome. Concretely, a medical diagnostic algorithm could detect disease with greater accuracy than a human doctor… make smaller errors on average. And yet it could systematically err on the side of false negatives (diagnosing no disease when disease exists) rather than false positives (diagnosing disease when no disease exists).

The interesting question is the extent to which deep human knowledge is an effective antidote to machine predictive error. Here is an interesting example.

In early 2022, Zillow lost $881 million and laid off a quarter of its staff after its home-flipping business led to catastrophic failures. On its surface, the plan seemed plausible: use the company’s vast data on home prices and a machine learning algorithm to predict which homes offered the greatest opportunity for purchase and resale. In essence Zillow said: if the machine learning algorithm predicts the home has a greater value than its current sales price, buy it.

For someone with a more sophisticated understanding of pricing (say, an undergraduate who had paid attention in their industrial organization or information economics class), they would have immediately recognized this as an example of the “winner’s curse”. Suppose a house has a true market value, but bidders at an auction all have different beliefs about that true value. The bidder who believes the house is worth the most will bid the most and win the house. Its only when they try to sell it that they realize the trap: they won the auction because they thought the house was more valuable than everyone else. And that’s a real problem if the whole idea is to flip a low valued house into a higher valued sale.

The idea that deep understanding of causal mechanisms can help us to catch mistakes is particularly important when we try to evaluate whether “knowledge” offered by a third party (whether it’s a podcaster or YouTuber or an LLM) is right or wrong.

For too many people and too many things we don’t know what we don’t know. What percentage of the human population that might employ an LLM or its predictions knows that planetary motion is governed by Newton’s laws? How would they know if the output is right or nonsense?

Deep Disciplinary Knowledge and GenAI are Complements

For someone who is equipped with strong frameworks for understanding the world, the weakness of LLMs (or machine learning more generally) becomes a strength. They become tools that can augment our abductions with rapid and semi-automated induction – quick evaluations of ideas. That could be a chatbot spitting text. It could be an LLM writing code for you to extract and rapidly analyze data.

But for an LLM to be valuable, you have to know what questions are worth asking, and how to ask them. And you also have to know when they are feeding you BS. And if you do, my goodness are these tools powerful!

A few weeks back I made my first try at “vibe coding”. I’ve never written any Python script and I wanted to extract some financial data. After an hour with Google Gemini I had 80% of what I needed. I couldn’t get the last 20% despite ten hours of investment and Gemini’s repeated insistence that it had succeeded (including swearing up and down that it had saved an output file it had not, in fact, saved.)

I think that if I knew even a little Python I could have corrected its code quickly. And luckily, I had just read some articles on the massive security breaches that agentic AI can create, so when it asked to install some executable on my machine I refused.

But the larger point was that I took a question I wanted to answer, identified the data, scraped and analyzed it (well, 80% of it), in about an hour. In the before times that would have been months of work to identify, train, and unleash a research assistant. Most likely, I would have decided it wasn’t worthwhile to do at all.

The labor augmentation versus automation part of this equation is that the LLM is not going to do any of this of its own accord. And even working with a human agent you have this problem: there are an infinite number of things one could investigate, questions you could ask with the help of an LLM, data patterns you could correlate. But the vast majority of these things are useless.

The value of a deep framework for an analysis lies in pointing you in the direction of what questions to ask, how to ask them, how to validate the answers you get, and how to communicate the results in a way that is both instructive to readers and informative about a deeper set of issues. Without this understanding, the output is actually worse than useless. It’s noise that drowns out real information.

Next Week

If you’ve stayed with us so far, congratulations (and thanks)! We’re going to try and pay off all your reading about AI with a final essay on lessons for students, faculty and universities...

“Finding Equilibrium” is coauthored by Jay Akridge, Professor of Agricultural Economics, Trustee Chair in Teaching and Learning Excellence, and Provost Emeritus at Purdue University and David Hummels, Distinguished Professor of Economics and Dean Emeritus at the Daniels School of Business at Purdue.

In the next two essays I’m going to get way out over my skis to try and tackle where all this might go next. Humans have been using tools for over 3 million years, some of which were merely useful, while others transformed entire civilizations. Is generative AI just another merely useful tool? Perhaps one that enables us to offload menial mental chores in the same way that the wheel enabled us to offload carrying heavy weight? Or, more troubling, might it supplant human cognitive labor across a wide swath of human endeavor? This turns on the question… what can we do that machines cannot?

I will focus on generative AI produced by large language models like ChatGPT, though I’ll comment in passing on other approaches. Why? LLMs are, by far, the most economically relevant AI tool in terms of investment, adoption, and for reasons I’ll describe below, have the potential to scale to wondrous (or devastating) impact.

I originally wrote this essay and the next for my undergraduate course on innovation and its impact on productivity, work, and society. Students have many questions about the technological story of our age – how will generative AI transform firms and work, and what should they be doing to prepare? I didn’t have great answers.

So, I did a lot of reading and thinking about artificial intelligence, about the nature of learning and reasoning, about how humans differ from the capabilities of generative AI. (These essays draw heavily on The Myth of Artificial Intelligence by Erik Larson, on Prediction Machines, and Power and Prediction, by Agrawal, Gans, and Goldfarb, and various writings by Gary Marcus.) Before I can explain anything to students, I need to first explain it to myself. This is my effort to explain these issues to myself.

To be clear, I’m an economist, not an AI expert. There will no doubt be experts in those fields who find some of this too simple, or wrong-headed or missing key parts of the story. (If so, comments are open, and be gentle!) My interest lies in understanding these things just well enough that I can think through the economic impact for firms and workers and the mission impact for universities.

Enough prologue, let’s go.

What are Large Language Models Doing?

Large Language Models like ChatGPT are trained on vast datasets, recognizing statistical regularities in text to generate highly probable “next token” responses. (A token can be a word or parts of a word or characters.) To understand why they are so important it is useful to start before the modern LLM era.

We use statistics to describe (organize, summarize) data, and to draw inferences based on that data. Inference includes both identifying relationships (correlations) between variables and using those correlations to make predictions. In my field of Economics, we are most interested in whether those correlations between variables represent causal linkages. For example, we want to understand not just if employment is correlated with minimum wage laws or tax rates. We want to understand whether changes in those policies caused the change in employment.

But in other instances, one might be more interested in the quality of prediction itself. If I’m getting dressed, I want to know how likely it is to rain, not why it’s going to rain.

Traditional linear regression tools provide both causal inference and prediction. Unfortunately, their predictive power is often quite low, especially when dealing with large amounts of data describing individual decision making. For example, these tools are not good at predicting whether someone who sees a web ad for a product will actually click through and purchase that product.

Advances in machine learning techniques put a premium on prediction; they seek to forecast some outcome as accurately as possible, even if we don’t understand anything about the relationships between variables that gave rise to that outcome. If I improve my prediction of whether someone will click on a web ad by just a small amount, I can better target those web ads and get much more bang for my advertising buck.

Those predictions and their consequences can also be automated. Google trains a machine learning model using your internet browsing behavior to predict your ad click through rates, and then automatically feeds you ads, no human intervention required. That is extremely valuable: Google generated a $238 billion in ad revenue in 2023. I previously wrote about the demand stimulation effect of new technologies that automate production. A quarter trillion dollars a year qualifies!

Now, there are many things we might want to predict. Not just an ad-click through rate or the weather, but stock prices, how much inventory a store needs to satisfy Black Friday shoppers, whether an aircraft engine will fail, or whether the dark figure in front of my self-driving car is a pedestrian or a shadow.

Efforts to classify images, to predict whether the collection of pixels on my screen is or is not a cat, launched the modern AI era. Researchers Krizhevsky, Sutskever, and Hinton from the University of Toronto used deep neural networks to dramatically improve image recognition. This work evolved to provide a foundation for transformer architectures that would become the backbone of large language models.

But stripping away all those details, it’s all just prediction. What many of us find astonishing about LLMs is what kinds of things you can predict and what power that puts in our hands. The ability to predict the next most likely language “token” enables the ability to produce credible essays and reports (to the lament of English professors everywhere), but also to provide rich description (organization, summarization) of data organized in the form of language, and in particular to write software code. Code is, after all, a particular sequence of characters arranged in a logical and therefore predictable manner.

Why is This Powerful ?

In their work on the economics of artificial intelligence, Prediction Machines, Agrawal, Gans and Goldfarb highlight the importance of prediction in decision making. Many of the choices we might make in life and in business depend on a forecast of what will happen in certain circumstances. Better predictions should lead to better choices and that creates value.

But prediction isn’t enough by itself. It has to be joined to the complementary skill of judgement. Put simply, I have to know what is worth predicting, and what I will do with the prediction I receive. When prediction becomes better and cheaper and faster, the value of that complementary judgement rises.

One of the ways this might play out in the job market relates to new versus experienced employees. New employees are often asked to do the kinds of things that LLMs, and their uncanny power to organize, summarize, and predict language, can now do. Beginning financial analysts, for example, might work their way through reams of company financial reports to spot patterns and opportunities. Associates in law firms might be tasked with producing lengthy legal analyses. What to do with these insights, what judgements to render, are left to more senior analysts and law partners.

To use the language of our first essay on AI, LLMs might automate the labor of new employees, and augment the labor of seasoned employees. How one becomes a seasoned employee if firms stop hiring junior employees is a different matter!

It is possible though that current LLM capabilities are just scratching the surface of where they might go. To see why, let’s talk about “The Bitter Lesson” an important essay by Rich Sutton, the father of reinforcement learning. Sutton writes that

“The biggest lesson that can be read from 70 years of AI research is that general methods that leverage computation are ultimately the most effective, and by a large margin…. Most AI research has been conducted as if the computation available to the agent were constant (in which case leveraging human knowledge would be one of the only ways to improve performance) but, over a slightly longer time than a typical research project, massively more computation inevitably becomes available...

“We have to learn the bitter lesson that building in how we think we think does not work in the long run. The bitter lesson is based on the historical observations that 1) AI researchers have often tried to build knowledge into their agents, 2) this always helps in the short term, and is personally satisfying to the researcher, but 3) in the long run it plateaus and even inhibits further progress, and 4) breakthrough progress eventually arrives by an opposing approach based on scaling computation by search and learning.”

Sutton is saying that having computer and cognitive scientists think hard about how people think and then trying to build that into AI doesn’t work. Or at least, it works poorly relative to what search and learning algorithms can accomplish because data and computing power are growing faster than human ingenuity.

AI companies have taken Sutton’s “Bitter Lesson” to heart, looking for breakthroughs through “scaling”. Each subsequent generation of LLMs involves larger training data sets, more parameters (weights) to describe the data, and more computation in massive data centers, to improve performance. Ironically, Sutton actually thinks LLMs are a dead-end. Not because of the reliance on computation but because they are engaged in “mimicking” language and not pursuing goal-based learning.

We wrote previously that the kinds of things one can currently do with LLM Ai’s is typically a small subset of the tasks a particular worker undertakes. And integrating the predictive powers of LLMs into a single person’s job, let alone a larger workflow, is not trivial.

Agrawal, Gans and Goldfarb’s second book on machine learning, Power and Prediction, highlights the difficulty in translating the promise of machine learning technologies into actual breakthrough products or profits. They point to the need to reorganize work and entire firms around AI’s predictive power before it will be truly transformative. There are good precedents for this: it took decades before manufacturing firms figured out how to use electricity (rather than steam driven power) to improve their productivity.

Where do the big AI companies think they are headed? In the short run, autonomous agents that can do a lot more than narrow predictive tasks. Agents that can replace an entire workflow, possibly even exercising their own judgement. In the long run, Artificial General Intelligence, in which an AI can replace *any* cognitive task a human can perform. If they get there, forget all this business about technology that automates versus augments cognitive labor. AGI will do it all.

Why You Should be Skeptical

Let’s go back to the fundamental way that LLM’s “reason”. Their success depends on recognizing what is common or likely within their training data, but they can falter when facing rare or contradictory situations. LLMs excel at pattern recognition but lack true understanding, which is to say they don’t engage in causal reasoning and so lack explanatory depth. This means they can also “hallucinate” information: offering statistically likely but patently untrue statements. It appears this is not a fixable bug, but instead a fundamental feature of their architecture.

This issue of whether some pattern occurs or occurs commonly enough in training data is sometimes linked to “scaling”. As in, if we can just give models enough data, and allow them enough parameters to predict the next most likely token in a sequence, they will do better and better at fitting the world.

But that doesn’t seem to be happening. Generative AI models have grown increasingly expensive to train, and yet, they keep making very simple mistakes. After OpenAI’s Sam Altman said that ChatGPT5 was going to be like having an advanced PhD student on your desktop it took only a few hours for users to find that it made the same basic “common sense” mistakes that earlier models had. It got simple math problems wrong. It cheated at chess.

The World Models Critique

A number of leading AI experts claim that the problems with LLMs cannot be solved by scaling because the challenge in more foundational than simply increasing the size of the training set. This has become known as the “world model” problem. (Here is a nice summary of the problem by Gary Marcus. AI pioneer and Turing Award Laureate Yann LeCun has a similar view and is evidently stepping out of his role in leading Meta’s AI efforts over the issue.)

The argument is that human cognition relies on a rich internal framework of sensory experiences, context specific knowledge, and causal relationships. This internal model allows humans to reason about unseen scenarios, predict outcomes, and understand nuanced contexts. To essentially create informed guesses about what is most likely to happen based on past experience.

Because current generative AI models lack these capabilities, they are prone to errors that would be easily caught by a human with a comprehensive understanding of the world. What does this human understanding look like?

Humans understand cause-and-effect relationships. We know that dropping an object will cause it to fall. If that object is a rubber ball it will bounce. If it is a glass it will break. If we walk on that broken glass without shoes, we’ll cut our feet badly. If we don’t bandage that cut, we’ll bleed all over the rug and maybe get an infection.

We start to understand these relationships even as small children and develop a common sense understanding of the world. We know that liquids will adapt their shape to fill a container but solids won’t. That solids cannot pass through walls.

We can track complex relationships over time and space, understanding how a story unfolds or how different objects relate in space. We know that something can disappear from view behind a wall but not have actually vanished from the world. (Babies figure out object permanence before they are a year old.) We know that a 20-year-old travel guide about a city is probably too out-of-date to rely on for restaurant recommendations or transit pricing information. But that a 20-year-old textbook on Shakespeare might still have excellent insights.

Our world models also encompass social and cultural norms that allow us to predict how people are likely to behave or respond in social settings. We understand that humans have goals and desires and these influence actions and decisions. That you can say things to friends and family you can’t say to strangers or to your boss.

Recent efforts to train generative AI have started to move away from language based training sets to video based training in order to induce a deeper understanding of the physical world. This could be highly useful in, e.g. programming robots to work in factories, but it is unclear at this time whether these efforts will reach the same heights of success that LLMs have in working with language. And this still doesn’t relate at all to human’s social, cultural, and temporal understanding.

Sophisticated World Models and the Limits of LLM Inference

When you first start working with generative AI models, they seem like magic. They also seem disturbingly close to human, and possibly even conscious, something known as the Eliza Effect. But then they start making mistakes that seem strikingly dumb. It can be easy to detect some LLM mistakes because they are so obviously nonsensical that even a small child would avoid making them.

It is perhaps harder to detect mistakes that require more sophisticated world models, especially for a human who is not themselves trained on that world model. Here’s one example.

In a recent test, scientists at Harvard and MIT wanted to understand if advanced LLMs were correctly reasoning about the world and creating their own sophisticated world model. They trained the LLMs on movements of planets within 10 million simulated solar systems and then asked for two things. One, predict planetary motions. Two, write down the model you derived to create that prediction (i.e. show the force vectors operating on those planets).

On the first, the models predicted planetary motions with high precision. On the second, they failed wildly, generating an expression that bore no resemblance to Newton’s Law of Gravitation. Most importantly, when looking at different solar systems in different galaxies the models generated completely different force laws.

If their predictions were accurate, does it matter whether they induced the correct fundamental law? Yes! Because if you were to change, e.g. the mass of Jupiter by 10%, their predictions would be wildly off. As it would be if you were to try to apply their force equations to another solar system with different sets of planets, which was evident when the models thought the laws of physics varied depending on the sample they examined. (Pro tip: Newton’s laws apply to planets in this galaxy, and in others, and to apples falling from trees.)

This is a classic case of “overfitting” the data. That is, generating explanations for patterns that fit a very specific data sample but are an inaccurate representation of underlying causal mechanisms. If one tries to apply these same explanations “out-of-sample” one can get things very wrong.

This failure makes Marcus and others extremely skeptical that LLMs will ever be able to do what humans can do. Which is to use a combination of reasoning methods and experience to arrive at fundamental and sophisticated world models that identify underlying casual mechanisms. When you understand the true causal mechanisms, your predictions apply out of sample. They apply to things we haven’t measured, or haven’t (yet) happened.

For a similar reason, Larson is extremely skeptical of the idea that that LLMs will ever be able to surpass human knowledge – develop new ideas as opposed to summarize old ones. The argument is that one cannot generate new understanding about the world without understanding the world at a causal level. If GenAI models cannot properly induce the causal mechanisms for human and physical phenomenon that we already understand, how can they induce things we don’t understand?

And if they cannot do that, more fanciful notions of a machine superintelligence that will replace the need for human innovation are straight out the window.

Equipping AI with World Models

An LLM can play chess, but if you provide it a sufficiently unusual situation, it will make illegal moves. And it won’t notice if you make illegal moves. The reason is that the “next token” predictive algorithms in LLMs fail to induce a robust “world model” of chess, even though it has trained on data that include both the rules of chess and countless historical games of chess. Which is to say, it can “read” a book on the rules of chess, but not “know” the common-sense implication… that anytime you play a game of chess you have to apply those rules.

About forty years ago I bought an inexpensive chess computer. It was preprogrammed with the rules of chess and a simple algorithm that assessed the value of different moves and possible outcomes multiple steps in advance. That pre-programming was the “world model”. I’m not a great chess player, so if I let it look more than a few moves ahead, that ancient computer consistently beat me.

But then 28 years ago, a more powerful version, IBM’s Deep Blue chess computer beat world champion Gary Kasparov. And 9 years ago, Google’s AlphaGO first beat a Go grand master, Lee Sedol. (Go is vastly more complicated than chess.)

The processing power available to, and investment in, LLMs has grown enormously relative to those older computers, but their chess (and Go) playing has gotten worse relative to dedicated tools. The difference is modern LLMs are not given the explicit world model to work with, and they can’t induce it on their own.

But maybe if you equip them first with a human understanding of how a specific game works (or how an accounting firm wants to identify fraudulent book keeping), they could be remarkably powerful.

Could you follow this logic to broaden AI capabilities more generally? Gary Marcus and Yann LeCun think the answer is yes. Rich Sutton’s Bitter Lesson would argue no, that search and learning based on computation, and not being pre-equipped with the scientist’s knowledge set, is the only way the field of AI has advanced.

Alas, equipping a computer with the rules of chess is trivial compared to equipping it with a comprehensive set of human knowledge across many domains. Increasingly it looks like companies are having modest success with pre-equipping models with specific rules to solve specific problems. But by their very nature these successes do not scale. Every use case is a unique, and costly to develop, which means it can only be successfully applied when there is a huge amount of demand for that use case.

Unless this changes, AI tools will remain “merely useful”. Like, say, a horse-drawn plow. Or Microsoft Excel. Powerful in particular contexts. Profitable. But not something that shatters the workforce or the role of universities. However, if scaling allows AI to generate world models on its own… that’s the point where civilization changes, and AI begins to replace large swathes of human cognition.

Me? I’m betting on the horse-drawn plow.

Next Week…and the Week After

We’ll give you a break next week to enjoy the Thanksgiving holiday! Essay four in this five-part AI series will be posted Friday, December 5. We will build on the ideas shared above to contrast what LLMs do with the unique nature of human cognition, and then use this to describe how LLMs increase the importance of experience and academic disciplines. As always, thanks for reading Finding Equilibrium!

“Finding Equilibrium” is coauthored by Jay Akridge, Professor of Agricultural Economics, Trustee Chair in Teaching and Learning Excellence, and Provost Emeritus at Purdue University and David Hummels, Distinguished Professor of Economics and Dean Emeritus at the Daniels School of Business at Purdue.

We promised then to take a deeper look. But digging into that question leads to another, even more important question: how is AI changing what new college graduates do in their first job/the work world they enter? The answer to this question will have fundamental implications for what universities teach and how they teach it. (David and I will explore those implications in our final post in this series.)

I acknowledge answering these questions with any degree of confidence at this point in time is exceedingly difficult: we are very early in employer adoption of AI; employers themselves are trying to figure out where AI adds value and where it does not; and all answers are transitory as the capabilities of AI continue to improve at a dramatic pace. (David and I have joked that we have to get this post up, because every day some new study/story comes out – in many cases contradicting the one that came the day before!)

With all those caveats, in this post I will take a brief look at what we know about AI adoption in the work world. I will spend more time on which jobs are most likely to be impacted by AI and the impact of AI on entry-level hiring.

Share

Get Beyond the Headlines…

There is no problem finding headline grabbing quotes about what AI will do to the work world and entry-level hiring. Wal-Mart’s CEO proclaimed “AI is going to change literally every job.” Ford’s CEO was even bolder claiming “AI is going to replace literally half of all white-collar workers in the US”. He didn’t say when…

Other headlines focus on how AI is ending entry-level positions as we know them. Of course, the firms investing billions to produce AI services have plenty to say about how AI will improve *everything*… There are headlines on the other side as well, claiming the job market impacts of AI are overblown.

Fortunately, scholars are digging deeply into important questions about AI’s impact on employment, productivity, wages, and just about every other labor issue you can think of – working to separate histrionics from reality. (If you want to go down a rabbit hole on the topic (like I did) take a look at the excellent collection of papers that Will Rinehart has assembled.)

Let’s get beyond the headlines and dig into what we know about what is going on.

How Many Employers Are Actually Using AI?

Despite all the hype, not that many - yet. The Business Trends and Outlook Survey of the Bureau of Census reported only about 10% of US firms answered ‘yes’ to the question ‘In the last two weeks did this business use AI in producing goods or services?’ But, as the chart below shows, the overall AI adoption rate (using AI to produce goods and services) by US business is growing rapidly.

Source: Claude.

Such aggregate numbers mask the variation in AI adoption rates across firm size and industries, as well as where and how intensively AI is being used by the firm.

Unpacking this, some recent evidence shows a slowdown in adoption of AI by firms with more than 250 employees and continued rapid adoption in smaller firms. AI adoption was about 5x higher for firms in the Information; Professional, Scientific, and Technical Services; Educational Services; and Finance and Insurance sectors relative to firms in the Transportation and Warehousing; Construction; and Agriculture sectors.

Such ‘business-level’ reporting likely undercounts actual usage: a MIT study found a thriving ‘shadow economy’ of AI users – 90% of the employees in their study reported using their personal AI accounts to help them do their jobs, often without firm knowledge or approval. But any improvements in individual productivity from AI use had not yet translated to improved firm ROI (maybe the employee had time for really long lunch breaks or started their weekend on Thursday?).

Bottom-line: While overall adoption is relatively low, firms are conducting (many) pilots and employees are exploring AI applications on their own. And, firms have aggressive expectations with respect to future AI adoption. This is not surprising: anecdotally, I have spoken with a number of mid-size/large employers in the past couple of months, and every one of them has some kind of group/task force/tiger team/etc. exploring AI applications in their firm – specific research applications, chatbots for customer service, back office finance and accounting applications, etc. So rapid expansion of AI deployment is a reasonable assumption going forward – but it is an assumption…

Is AI Currently Paying Off?

Depends on who you ask. AI is being touted as a general-purpose technology (GPT), a technology that can be applied very, very broadly across the economy. Examples of general-purpose technologies would be electricity, the internal combustion engine, and more recently, computers and the internet.

Firms hope AI can be used widely within their operations to help them improve productivity by increasing output, improving quality, lowering costs, etc. – and, as David pointed out last week, perhaps opening up entirely new business opportunities that are not on anyone’s radar today.

Well, that’s the dream, but we are not there yet - the overall productivity impacts of AI barely register. A study by the St. Louis Fed asked workers how many days per week and minutes per day they were using generative AI in their jobs – and how much time AI was saving them.

Use was limited - only about 28% of workers reported using AI in their jobs (November 2024). Not surprisingly though, the more a worker used AI, the higher their reported time savings. Almost 53% of those who reported using AI every day (9% of all workers) reported at least 3 hours of time savings while about 1/3 of the workers who used AI at least once a week, but not every day (14% of all workers), reported similar time saved.

Overall, the study found that each hour of AI use by workers is on average 33% more productive. But, given the limited use, the average time savings across all AI users was about 5.4% of a work week - 2.2 hours per week.

There are of course big differences across industries with those in information; finance; and professional, scientific, and technical industries using AI much more intensively and reaping more time savings relative to occupations such as accommodations and food service, and transportation. (The industry time savings figures shown in the chart below are lower because they include all workers in the industry - both AI users and non-users.)

Source: St. Louis Fed.

These findings likely reflect the point David made last week: most jobs are a big collection of many different kinds of tasks, and generative AI is only helpful in performing some of those tasks. But the 33% increase in hourly productivity reported for AI use is something to watch as AI capabilities evolve and workers find ways to use AI to perform additional tasks.

Given the varied findings I have shared to this point, it won’t be a surprise that overall corporate enthusiasm for AI is all over the map. MIT reported in July only about 5% of firms embracing AI are generating a return on their investment – 95% reported zero return at this point in time. Part of the MIT story is larger firms are struggling to scale their investments. Part of their story is that AI investments have been made in very visible top-line functions such as sales and marketing where productivity gains have been marginal, and not in hidden but important back-office functions such as legal, finance, and purchasing where substantial productivity gains are possible.

Painting a radically different picture, Wharton just released a study that found nearly 75% of the firms they surveyed were seeing a *positive* ROI, with mid-size and smaller firms more likely to report a positive return more quickly. They argue firms are now moving quickly beyond pilots and embedding AI in existing workflows such as code writing, HR functions such as recruiting and onboarding, and legal contact generation, as well as finding broad AI applications for tasks such as data analysis, document summarization, and document editing/writing.

These kinds of highly contradictory findings are the order of the day when it comes to AI! If you to want to better understand why, take a look at Alberto Romero’s post “Why You Can’t Trust Most AI Studies”. Given we are very early in AI’s adoption phase, it is difficult to draw general conclusions and what a study finds depends on who they survey, what they ask, and the success metrics employed (and perhaps who funded the study…). Of course, vanilla, middle of the road findings (‘AI is trying to find a path’) don’t get headlines…’AI the all-powerful’ and ‘AI the giant tech bubble’ do.

So, What IS Going on Inside Firms?

Broad integration of a new technology is hard. lt takes time for workflows to be changed, eliminated, and/or recreated to capture the potential benefits of AI. People have to be trained on the new technology, new processes, new roles. The AI tools must be developed and adapted to deliver something superior to whatever was being done formerly.

One example of such ‘growing pains’ is a dramatic increase in ‘workslop’. Workslop is defined as “AI generated work content that masquerades as good work but lacks substance to meaningfully advance a given task.” The idea here is an employee puts AI to work, generating a slick report, a colorful presentation, a well-worded memo – none of which say much (or anything). And, while the work product looks good on the surface, there is no/little substance, forcing someone else to spend even MORE time making sense/correcting the work product.

The authors of the workslop study reported 40% of employees surveyed had received workslop in the last month, requiring an average of 2 hours to address in each instance. (That’s about the same figure as the average time savings reported for AI use above!) (I am reminded of the old Calvin and Hobbes cartoon where Calvin is asked to write a report on bats. His report is one sentence: Bats are bugs. Nonetheless, he argues for an ‘A’, exclaiming “that my report is in a professional, clear plastic binder!’’.)

There is nothing unique about the current state of AI adoption. Many technologies actually lessen productivity before the full benefits are realized.

A Brief Interlude: Where’s the Beef?

Now, you might say…if few firms are using AI and returns are questionable, why are these AI firms spending so much money? US investment in data centers this year is expected to reach $400 billion! Are these guys just idiots?

Think about this math. Suppose the net effect of AI replacing some jobs and improving the productivity of others amounts to 1% of GDP – just a ‘measly’ 1%. That’s worth $300 billion this year, and every year after…so about $6 trillion in the US economy and maybe $20 trillion world-wide. And, that does not account for ANY of the kind of breakthroughs/new business opportunities AI might create.

AI does not have to replace a lot of jobs to be incredibly valuable! And, Amazon just announced they would be cutting 14,000 jobs (1% of their workforce), UPS 14,000, Target 1,800 (8%)…all white-collar jobs…

Now, back to our story…

Subscribe now

Which Jobs are Exposed to AI?

Scholars have developed a number of different measures of ‘AI exposure’ – basically how likely a job is to be affected in some way by AI. That effect could be positive and enhance the productivity of the person doing the job. Or, the effect could be negative replacing some or all of the tasks performed by someone holding the job. The first effect has come to be known as augmentation and the second automation.

These AI exposure measures typically link tasks that AI can perform to a set of job skills, then link those job skills to the skills required in specific occupations. To illustrate, let’s assume writing reports (a task AI can perform) could be seamlessly pulled out of a job’s workflow; if one occupation consists of spending 30% of the time writing reports and another spends 10% writing reports, the first occupation is more AI exposed.

Such occupation data can then be rolled up to the industry level to assess industry exposure to AI.

While many of these AI exposure classification studies are based on a set of assumptions about the link between AI capabilities and specific job tasks, an interesting study (from Microsoft Research) used 200,000 anonymized Co-Pilot (2024) conversations to look at ‘user goals’ (seeking AI assistance with a task such as developing content) and ‘AI action’ (tasks the AI actually performed such as responding to customer inquiries) to get at AI exposure. This distinction starts to help sort out tasks AI augments (user goals) and tasks AI automates (AI action).

These user goals and AI tasks from the 200,000 conversations were mapped against a set of 332 (O*NET) work activities which were then rolled up to specific occupations. User thumbs up/thumbs down feedback on the Co-Pilot conversation was used as a measure of AI effectiveness.

The most frequently AI-impacted work activities are shown in the figure below with the X-axis showing the frequency each work activity was reported as a user goal or an AI action. Summarizing this figure, workers are using AI to gather, process and disseminate information (user goals) and AI is gathering, explaining, and communicating information to the user (AI tasks). To me, these results suggest AI is more likely being used currently to enhance the productivity of workers (augmentation) relative to replacing what they do (automation).

Source: Microsoft Research.

Continuing with the Microsoft study (their results are pretty consistent with other attempts to classify occupations by AI exposure), the table below shows occupational groups by their AI Applicability Score (higher score, more likely to be impacted by AI). To overgeneralize, jobs that involve basic communications (writing), mathematical, analytic, and coding skills are (currently) most exposed to AI. Jobs that involve physical labor and/or direct human contact tend to be least exposed to AI.

Source: Microsoft Research.

How does all this relate to college degrees? Indeed looked recently at the relationship between employer degree requirements in position announcements and exposure of those positions to AI. ‘Exposure’ as defined by Indeed could be automation, or it could be augmentation. In general, the more likely an employer was to specify a degree requirement in a job announcement, the higher the level of AI exposure for that job – an important message for universities!

Source: Indeed Hiring Lab

What is the Impact of AI on New Graduate Employment?

The job market for new college graduates HAS continued to soften – with unemployment for 20-24 year olds holding a bachelor’s degree ticking up to 9.3% in August – the highest rate since just after the pandemic. As I mentioned up front in this post, some blame this increase in unemployment on AI.

One argument here is that many of the tasks that a new college graduate performs are somewhat basic, requiring codified, explicit knowledge such as information gathering, basic data analysis, report writing, presentation development, etc. – the kind of stuff AI (even now) is pretty good at. So, fuel for the argument, who needs new hires anymore?

Source: St. Louis Fed.

While different studies come to different conclusions on this point, in aggregate, it does not look like AI is to blame. David and I discussed this in an earlier post – many of the current trends we are seeing in the job market for new college graduates started well before ChatGPT was a thing. And, this recent downturn in hiring is taking place during a period of policy uncertainty (tariffs), a sharp drop in consumer confidence, and after a huge surge in hiring in some of the affected fields.

However, while the *broader* impacts of AI on new college graduate hiring in general are limited, evidence is mounting that there is some impact on the most AI exposed jobs. Using payroll data from ADP, the payroll processing company, economists from Stanford (and NBER) showed early-career workers in the most AI-exposed jobs have seen a 13 percent relative decline in employment!

Source: Stanford.

Another study (Maasoum and Lichtinger ) also found that while the effect of AI on younger workers overall was negligible, employment of recent college graduates at firms adopting AI was lower (reduced hiring, not separations). Basically, these AI-adopting firms are placing bets that AI will improve productivity allowing them to scale back hiring currently for roles they predict will be automated/augmented in the future. We will see if these productivity gains materialize…or if they will need to come back to campus career fairs…

Summary: Where are We?

• Firm adoption of AI at this moment in time is relatively low. But, it varies by industry, and firms are very much in the pilot/experimentation stage – with ambitious expansion plans.

• AI is being used informally in the work world at much higher rates than formal adoption and such informal adoption will help firms sort out where AI creates value – and where it doesn’t.

• While actual AI use by workers is modest and time savings are minimal, some strong productivity gains have been reported when and where AI is being used.

• It is too early to pass judgement on AI’s payoff for firms. Interesting inefficiencies have been created (workslop) as firms attempt to deploy AI within existing workflows. Broad evidence of positive returns isn’t there, but specific wins have been documented.

• The most AI exposed jobs at this point in time are ones that utilize codified, explicit knowledge and don’t involve as much direct human contact – jobs that require a college degree.

• AI has not had a dramatic effect on the hiring of new college graduates in the aggregate, but there has been a pullback in hiring with some of the most AI exposed jobs/jobs with firms adopting AI.

• Plenty of employers are signaling their bets on AI by announcing workforce reductions/slowing hiring. So, where we are now is much less important than where we may be going.

• Whether or not we blame the increase in recent college graduate unemployment on AI, class of 2026 graduates face a softer job market and likely need additional university support with their job search…

Next Week

Zealots say AI will take over the world. Cynics say we have an AI bubble that is about to burst. We are in a time of massive experimentation with a technology that is evolving literally daily.

Will AI tools ultimately amount to souped-up calculators to which we offload difficult but menial cognitive chores, or will they act as autonomous “agents” or even superhuman intelligences capable of their own breakthroughs? David will contrast human methods of reasoning with the algorithms employed in large language models next week. We hope you tune in for the third episode in this five part series on AI.

Research assistance provided by Marley Heritier.

“Finding Equilibrium” is coauthored by Jay Akridge, Professor of Agricultural Economics, Trustee Chair in Teaching and Learning Excellence, and Provost Emeritus at Purdue University and David Hummels, Distinguished Professor of Economics and Dean Emeritus at the Daniels School of Business at Purdue.

Universities and faculty are struggling to respond to the influx of generative AI tools and their impact on our research and our teaching. Should we articulate campus-wide AI policies or leave it to faculty? Should we ban AI, or teach it? Do we give up on take home essays and make them handwrite in class?

These are important questions. But they are less important than the question of how AI might reshape our students’ career opportunities and what that implies for the entire enterprise of higher education. We have put together a five (!!) part series to help our colleagues on faculty and in administration understand the underlying issues. Then they can get back to the question of how hard it is to decipher handwritten in-class essays…

Here is what you will read over the next five weeks. (And if you find our perspective helpful, please share with colleagues and invite them to subscribe!)

Share

1. What Automation and Offshoring Tell us About The Labor Market Effects of Artificial Intelligence

We have decades of experience with the automation of farm and factory, and the offshoring of production to low wage countries. What can we learn from these historical analogues to AI about labor substitution, augmentation, and demand stimulation?

2. AI Adoption and Effects on Entry-level Hiring and Jobs: The Evidence So Far

It has been three-years since ChatGPT 3.5 exploded on the scene, and AI firms are making truly massive investments to capitalize. What do we know about AI adoption by firms, changes in entry-level jobs and hiring patterns, and the productivity of workers and firms who adopt AI?

3. How is Human Reasoning Different from Predictive AI Tools?

Will AI tools ultimately amount to souped-up calculators to which we offload difficult but menial cognitive chores, or will they act as autonomous “agents” or even superhuman intelligences capable of their own breakthroughs?

4: Experience and the Role of Academic Disciplines

We explore why AI can’t replace lived experience and understanding the causal analytical frameworks offered by academic disciplines, but makes these tools even more powerful for performance and advancing human knowledge.

5. Lessons for University Strategy

We don’t know where all of this will ultimately land, but universities can’t wait for some steady state to emerge. We pull our previous four posts together in a set of ideas for universities to consider as they develop their ‘AI strategy’ during this time of rapid change.

Let’s get started…

What Automation and Offshoring Tell us About The Labor Market Effects of Artificial Intelligence

While it is unclear precisely how artificial intelligence will affect the economy, much of the conversation has focused on the possibility that these AI tools will begin to replace large numbers of workers. This is not the first time we’ve heard such concerns – automation of work on the farm and in the factory, and the offshoring of manufacturing have led to similar replacement effects in decades past. What can we learn from these episodes?

In 1850, roughly 2/3 of the US workforce was employed in agricultural production. But then, better tools, mechanization, and improvements in irrigation, seed, fertilizers, and pesticides sharply increased farm yields. Between 1950 and 2021, farm productivity improved by a factor of 3, and the labor employed in agriculture fell dramatically, reaching just 1.2 percent of the US workforce.

Starting in the 1980s manufacturing firms began to substitute industrial robots for labor, and beginning around 2000, offshoring of production accelerated that trend as companies moved manufacturing to other countries. Even though US manufacturing output rose 80 percent, labor’s share of manufacturing value-added fell from 68 to 52 percent, and the number of workers employed in US manufacturing fell from 19.5 million in 1979 (20% of the workforce) to 12.8 million (8% of the workforce) today.

Labor Effects: Substitutes, Complements, and Demand Stimulation

These stories seem straightforward – automation and offshoring take jobs, and we might expect AI will do the same…only for white collar workers this time. But the overall trends noted above mask three competing effects. The first is labor substitution. When firms and farms substitute more efficient capital for labor, they reduce the amount of labor needed for each bushel of wheat or car produced.

The second effect is labor complementarity. Some technologies augment the productivity of certain kinds of labor. Factory automation replaced labor used in assembly but created demand for more educated workers who design and implement automated systems. Computers replaced typists and manual bookkeepers, but created demand for anyone who could design or program computers, or even just use the power of computers to accelerate or improve the quality their own work. Offshoring allowed foreign workers to replace US workers with low education levels, but augmented demand for US workers with high education levels.

The third effect is demand stimulation. In agriculture, productivity growth has lowered the price of agricultural output by 50 percent since the 1950s. The (quality adjusted) price of consumer electronics has fallen 95 percent since 1997. Lower prices lead consumers to buy more of the product, and can actually increase the total amount of labor employed. (For example, suppose automation enables us to reduce labor employed in making a car from 100 hours to 80, but car sales double from 50 to 100. Total labor employed goes from 5000 hours to 8000 hours.)

The demand stimulation effect need not work only on prices for a fixed set of products. Technological change and offshoring also lead firms to improve the quality of products, offering better features or performance, or creating entirely new items. These quality and product variety effects can stimulate demand more dramatically than simply making something cheaper.

Labor substitution is the main focus of AI-commentators. They ask: how many jobs will be lost as AI performs work now done by people? They should ask two more questions. Which types of labor will AI augment, and how much will AI-enabled firms and industries stimulate demand if prices drop or new products are put on offer?

The labor substitution/labor augmentation effects are really interesting because they get at some fundamental questions. Like, what are humans capable of that AI is not? And, what would make it difficult for firms to engage in wholesale adoption of AI to replace large chunks of their labor force? And what the heck should universities and their students do about all this? But first, let’s talk about the historical evidence on demand stimulation.

Historical Evidence on Demand Stimulation

Demand stimulation asks: how responsive is consumer demand to changes in prices, product quality and new variety? Let’s focus on demand for a simple product like wheat, or a pair of shoes. If only some firms within an industry adopt a new technology that lowers labor costs for those firms, employment will paradoxically increase in those firms and decrease in competitors. But an industry-wide shift toward labor saving technology is likely to reduce overall employment in that industry.[1]

In the case of agriculture, people responded to falling prices by eating more, eating differently (more animal protein), and using agricultural products for industrial applications (biofuels). But despite our best efforts and growing waistlines, we didn’t eat enough or find enough industrial markets to make up for tripled farm productivity. Labor employed in agriculture fell dramatically.

As for footwear, here’s a fascinating graphic from the Washington Post, which shows what happened with employment in shoe making. First, a wave of factory mechanization replaced traditional cobblers, then offshoring moved that shoe production overseas. We’ve all got a lot more shoes in our closet than people had in 1850… but not enough to employ 2.75% of the US workforce. (That would be 4.5 million shoemakers today…)

Source: Washington Post.

But focusing on wheat, or shoes, misses the most interesting parts of what happens in the modern economy. Increasingly we find that technological change and more efficient production enables an explosion of new products and services we didn’t even know we wanted. In 1943, the Chairman of IBM, Thomas Watson, offered the hysterically inapt prediction, “I think there is a world market for maybe five computers.” Let’s think about just one company.

Apple began life in 1976 with Steves Jobs and Wozniak soldering motherboards in their garage. By 1998, Apple employed 6658 people and had a market capitalization of around $2 billion. Today it is worth 2000 times that amount, just under $4 trillion, directly employs 164,000 people worldwide, contracts with many more through foreign suppliers, and claims to “support” another 2 million US jobs through suppliers and app store creators. That is demand stimulation on steroids.

What happened? Apple’s employment growth really takes off in 2007, with the launch of the iPhone. iPhone sales now directly generate more than half of Apple’s revenues, and if you include all the ancillary services and devices that iPhones generate and support (think: App store sales, Apple Watch), it’s more like 80%.

Source: https://stockanalysis.com/stocks/aapl/employees/

Its notable that Apple doesn’t directly manufacture any of that stuff in the US. That explosive growth corresponds to Apple’s ability to combine its technological design and marketing prowess with offshoring high-quality manufacturing to low cost Asian countries. This allowed them to hit a price point with their iPhone that made phone ownership possible for a very large market! But it also facilitated a massive expansion of auxiliary applications for services we didn’t even know we needed.

The larger point is that its simply wrong to think about production as a zero-sum contest between capital and labor, or between labor in the US versus labor in Asia, or between labor provided by people versus “labor” provided by AI agents for a fixed set of goods. The creatively destructive powers of technological change are most evident, not when they are lowering costs, but when they are expanding the set of what we can do and what we make available for people to consume.

Disaggregating Tasks

In many ways, incorporating AI into the workforce is much closer to offshoring production than it is to mechanizing agriculture or automating factory floors. We are asking a machine intelligence that we do not fully understand or control to handle for us a subset of tasks that we used to perform ourselves. Let’s build out the analogy and draw on decades of experience in offshoring production to provide some useful insights.

Modern production involves many distinct stages (aka tasks). For manufacturing that means R&D, product design and testing, parts and components manufacture, assembly, and marketing and distribution, all undergirded by a set of “back-office” services including IT, HR, and finance. Service industries have similar stages, but without the physical artifacts of goods or their subcomponents.

Because each of these stages is produced using different inputs (R&D needs scientists and engineers; assembly needs manual labor), firms can reduce expenses by moving production of some stages/tasks to locations that provide those inputs less expensively. Back-office IT services get offshored to India. Manufacturing assembly and then later most of the manufacturing value chain are offshored to China or other locations around Asia. Each time this happens the labor share of output for the offshoring firm falls, but costs also fall and as a result, sales increase. The net effect is often to increase employment in that firm by stimulating demand.

What Caused Offshoring to Explode?

Not all firms and not all processes are amenable to offshoring. And even though the logic of international specialization based on stages of production has always made sense, offshoring didn’t really take off until the advent of significant changes in the technological and policy environment from the mid-1990s onward.

Keep in mind that when a firm separates production across locations, it creates a more complicated workflow. Even if a foreign location can produce at lower cost, information and physical products have to flow from one location to another. (And repeatedly! The North American automobile value chain involves parts and components crossing international boundaries 7-8 times.) The explosion in offshoring didn’t happen until there were sharp reductions in the costs of moving goods and information across borders and oceans.

Transportation costs fell due to containerization of ocean cargo and jet engines, and airplanes bridged both distance and time with profound implications for reorganized production. Political barriers to segmenting production fell, including lower tariffs, and heightened protections for intellectual property and foreign investment. Information technology made it possible to directly offshore back-office services, while also providing a dramatically improved flow of information across segmented production locations. Firms could communicate design specifications, track the flow of inputs, provide instantaneous information about defects or stockouts.

Verification, that is, ensuring the quality and reliability of foreign production is critical in this process. It’s no bargain to purchase low cost inputs from abroad if they arrive late, or in the wrong amounts, or suffering quality defects. Planned cost savings ultimately lower profits if quality or delivery problems shut down a production line or lose loyal customers.

Lessons for AI in the Workforce

What can we learn from the offshoring experiment about the use of AI in the workforce?

First, think in terms of tasks, not in terms of overall employment or even integrated occupations as they have been traditionally defined. Apple doesn’t think success means owning every stage of an iPhone’s creation. It grew 2000 times more valuable when it stopped trying to do everything; it focuses on design, software, marketing, and distribution and lets Asian suppliers handle the manufacturing.

We can reframe work by asking: how does a worker divide their day or week across many distinct tasks? You might try this exercise. Get out some paper and write down every distinct thing you do during the day, and how much time you spend doing it. Then at the end of the day, revisit that, and add all the things you forgot. If you’re like most people, you don’t do one thing. You do 50.

Second, of all the tasks you do, which of these many tasks might be subject to automation or augmentation by AI tools? This is a deeply interesting topic so we will tackle it at length in the next three essays. The really short answer is that AI tools are “only” good at prediction. But what we’re starting to learn is how many things can be predicted – not just the weather, but words – and how powerful that might be.

Third, individual tasks are not valuable by themselves, any more than Apple’s design or manufacturing is valuable in isolation. It’s the efficient combination of tasks to deliver a final product or service that is valuable to firms. Even though these AI tools appear miraculous at times, their use-cases are often much narrower than the ways we traditionally define someone’s work day.

As an example, OpenAI has created a new set of benchmarks aimed at measuring its ability to perform “real world” tasks that we might require of various occupations. What struck me in reading this was how narrow the task set was, how little of a person’s day might be absorbed in doing the tasks OpenAI wanted to automate for you. In our next essay we will provide some evidence on just how much people are using generative AI tools currently and how much time it saves them. Answer: not that much. Are you really going to lay someone off because a tool can do 5% of their job for them?

Fourth, separating and then later recombining distinct tasks is not trivial. Offshoring of production didn’t take off until it became much cheaper and easier for information and goods to flow across borders. Even if AI can automate or augment particular tasks, what is the workflow between those tasks and what are the barriers to their re-integration if we offload some but not all of these tasks to AI tools?

Barriers to Integrating AI Tools into the Workflow

The transformer architecture that powers modern generative AI tools predated by 5 years the explosive launch of ChatGPT in late 2022. But it was the chatbot interface that made the tool useful to many by lowering the costs of interacting with it. As we go deeper into integration with AI tools, there are several barriers to interaction and re-integration that remain significant and perhaps insurmountable: regulation, verification, and security.

As US firms are currently learning, the best offshoring plan will fall apart if government regulation and taxation intervenes. The big AI integration risks here are: copyright; government regulation of privacy and confidentiality, particularly around medical, financial, and personnel data, but also related to child safety; and regulation of physical objects such as driverless cars that embody AI-use.

Verification and minimizing defects are a major problem with generative AI. Modern quality control in manufacturing only tolerates exceptionally rare minor defects: a “six-sigma” standard implies tolerating something like 3-4 defects out of every one million parts. Without it, offshoring production to suppliers who operate at a distance and with long lead times would be infeasible.

The hallucination rate for the best-performing generative AI on the easiest problems is something like 1000 times worse than that. Many experts (including evidently OpenAI’s scientists…) think hallucination is fundamentally unfixable and inherent to the architecture. That has financial consequences for adopting firms. And becomes a source of great embarrassment and possibly legal sanctions for other professionals such as attorneys, and even judges. Worse, the more we hand complex and detailed work to AI agents to perform, the less we know about the work it is doing, and the harder it is for us to detect these defects.

Finally, a potential solution to the task handoff problem is to allow AI agent to integrate more deeply with our workflow and act with greater autonomy. If you check every step of their work, you have lost all of the time savings you had hoped for (and possibly more). But allowing autonomy means they are doing things you don’t understand and you don’t control, and that may have extreme security vulnerabilities. (Go read this 2023 NIST report on adversarial machine learning, or this update on AI attack vectors and be very afraid.)

Summary… And What’s Coming Next Week

My hypothesis is that in the near term most firms will struggle to significantly reduce labor inputs, even in cases where AI tools are demonstrably better and cheaper at some tasks. The reason is that too much of workers’ task sets are not amenable to AI disruption, and the handoffs between tasks sufficiently difficult to navigate, that they will offset significant savings. Fully exploiting AI tools means completely rethinking task sets and workflows.

There may be some exceptions to this. Not all jobs are complex, some consist of doing the same “routine” task over and over. As it happens, those are precisely the jobs that were offshored over the last few decades. And likely the ones that will disappear in an AI-enabled firm.

The more powerful and interesting case is that augmenting labor with AI tools will lead to the creation of entirely new products and services, which not only create value for firms, it causes the most successful among them to actually expand employment.

These predictions are based on historical analogies. But things are moving fast! So what IS going on currently with AI adoption and labor substitution/augmentation? Importantly for universities, how is it impacting entry-level hiring and the work world new graduates are moving into? Jay digs into those questions next week.

Finding Equilibrium” is coauthored by Jay Akridge, Professor of Agricultural Economics, Trustee Chair in Teaching and Learning Excellence, and Provost Emeritus at Purdue University and David Hummels, Distinguished Professor of Economics and Dean Emeritus at the Daniels School of Business at Purdue.

Subscribe now

[1] Demand for specific firms within most industries tends to be highly responsive to changes in prices, whereas demand for an industry as a whole tends to be much less responsive. For example, suppose we’ve got two grocery stores in town, Kroger and Meijer, and Kroger cuts its prices by 10%. Kroger would likely attract so many customers from its rival that its revenues would increase, while Meijer’s revenues decreased. That is a within-industry reallocation. But if both stores cut their prices by 10%, people would use some of that savings for other things (eating in restaurants, buying clothes or electronics) and total grocery revenues for the two stores would likely fall. Related, the research literature shows clearly that when a specific firm offshores production more than a typical firm in that industry, its productivity rises and sales increase relative to those other firms. The reason is that most firms face, within their industry, elastic demand for their products – quantities increase faster than prices drop. But that won’t happen if *every* firm in that industry drops prices by the same amount.

There must be some standard ‘job crisis story’ template out there (maybe AI generated). It seems almost every story starts with comments from a graduate who hasn’t been able to find a job after 100s or 1000s of applications and is questioning the value of their college degree. Then some data are cited from the Federal Reserve showing the unemployment rate for college-educated 22-27 year-olds is higher than the overall unemployment rate, followed by a few comments from employers who aren’t hiring as many new graduates, a hint that AI is responsible for the problem, and then closing with a despondent comment from the unemployed graduate. Sigh…

Now we are not heartless: we are very sympathetic with any person who wants a job and can’t find one – and especially a person who spent several years and a lot of money earning a college degree. But what is really going on here and why? Are our new graduates facing an employment crisis? Is a college degree no longer valuable in today’s work world? Let’s take a look.

The Fed Data

It is true that the unemployment rate for college-educated 22-27 year-olds is currently higher than the unemployment rate for all workers. That is unusual: since at least 1990 recent college graduates have had lower unemployment rates than the average worker.

Source: Federal Reserve.

A chart from Derek Thompson (below) shows vividly what has been going on with new college grad unemployment. He calculates the difference in unemployment rates between all workers and recent college grads (a value of 1.5 means that the unemployment rate for all workers was 1.5 percentage points higher than that for recent college grads). Starting in 1990 and through the start of the COVID pandemic, the overall unemployment rate was higher than that for new college graduates. About 2019, the gap turned negative, with recent college grads having a higher unemployment rate than the average of all workers.

Source: Derek Thompson

But a few things are really notable here. First, the gap in unemployment rates (i.e., the advantage recent grads had relative to all workers) has been trending down since 2012. So, it might be sensible in identifying the source of this change to look for causes that go back more than a decade.

Second, as the first chart shows, college-educated 22-27 year-olds are still employed at higher rates than the average for all 22-27 year-olds. So maybe the job market doesn’t like *any* young people…

Third, it might be more instructive to compare recent college grads to all college graduates. We do that in the chart below. While unemployment among 22-27 year-olds with a college degree is always higher than that for older college graduates, this gap has been widening steadily since the turn of the century, growing from under 1 percentage point to almost 3 percentage points recently. The story here: employment prospects for new college grads have been softening relative to all college grads for a while. (You can see the havoc COVID-19 wreaked on the labor market in all these charts...)

Source: Author calculations from U.S. Census Bureau and U.S. Bureau of Labor Statistics, Current Population Survey.

Bottom-line: the current market for new college graduates is clearly weaker than that in recent years. Let’s dig deeper…

Where are the Problem Spots?

You may be surprised to learn that among the 9 majors where unemployment for 22-27 year olds is the highest, 5 are STEM fields: physics (7.8%), computer engineering (7.5%), computer science (6.1%), chemistry (6.1%), and information systems and management (5.6%). (The other 4 majors with high unemployment rates are: anthropology (9.2%), commercial art and graphic design (7.2%), fine arts (7.0%), and sociology (6.7%).)

Recent male graduates are unemployed (6%) at higher rates than female graduates (3.5%). Much of the gap looks driven by gender differences in career paths. Hiring has remained strong in the health care fields dominated by women, while the tech sector and manufacturing, typically male dominated, have been much weaker.

Subscribe now

What is Going On – The Transitory Argument

Why the soft job market? And, how long will it last? In truth, no one knows. However, plenty of hypotheses have been offered up – some are transitory arguments, some say it is a structural change in the market for college graduates.

We certainly buy some of the transitory arguments. Hiring spiked after COVID-19 as workers were job hopping like crazy and employers were looking to fill open positions/grab talent where they could. The tech sector was on a roll and was especially aggressive. All told, it was a great time to be a new college graduate!

Now, we are looking at a period of adjustment: the tech sector (132,000 layoffs so far in 2025, 238,000 in 2024) and the federal government (275,000 jobs cut to date in 2025) have been shedding jobs so there are plenty of people with work experience looking for positions. Many of these layoffs were last-in first-out, so younger workers bore the brunt.

The economic policies of the current administration have created enormous uncertainty and employers (especially in manufacturing, professional and business services, and the wholesale trade) are reining in hiring as they ‘wait and see’ how the impact of tariffs, the One Big Beautiful Bill, Federal Reserve battles, etc. play out.

The consequence of the recent policy environment is seen clearly in the NACE survey of employer hiring plans. Their fall 2024 survey projected a robust rebound in jobs for new college students in 2025. But then their Spring 2025 survey revised those estimates downward dramatically – from a 7.3% increase in available jobs for Spring 2025 graduates to a 0.6% increase – most likely a result of all the uncertainty employers were facing this Spring.

The post-COVID ‘great resignation’ now means something different: instead of resigning en masse, employees are resigned to their current job – labor market data show job openings, hires, and separations have all been quite stable over the past year, with separations (especially “quits”) running below pre-pandemic levels. (There is even a term for this business of hanging on to your current position: ‘job hugging’.) With fewer people exiting roles, fewer openings arise from turnover, and the steady number of hires tends to favor experienced workers – tightening the pipeline for fresh college graduates.

Having said all this, it may be helpful to put some of these numbers in perspective…currently, universities supply about 2 million new bachelor’s degree graduates annually. When the unemployment rate for this group increased from 4.2% in June 2023 to 4.8% in June 2025, it means the number unemployed from each graduating class rose by an additional 12,000 people. That’s not great news for those 12,000, but that number kind of pales relative to the 645,000 jobs cuts in tech and the federal government over the past couple of years…

What is Going On – The Structural Argument

(One) structural argument goes basically like this: the number of college graduates has grown over time, but the number of jobs requiring a college degree has not grown as fast. Even non-economists know what happens when supply exceeds demand. One piece of evidence here is that the college wage premium (while still very substantial) has plateaued and declined slightly since the early 2010s – and with the number of college graduates on the market growing faster than the number of jobs requiring a college education, unemployment has gone up.

Exploring this point further, what about those high unemployment rates for computer science and computer engineering graduates? Do we have an even more acute supply and demand problem here?

There is no question that higher education responded to the call for more ‘coders’ over the past 15 years. More than 600,000 students are studying computer science in the US today. As a result, the number of new BS degrees in computer and information science has exploded – from about 40,000 in 2008 to over 100,000 in 2021.

Source: WSJ

Looking at jobs in computing (broadly defined): while the number of new graduates has surged, employment opportunities have not kept pace. This can be seen in the chart below by comparing the annual growth in bachelor’s degree holders (201%) to the growth rate in employment among all workers in computing fields (57%). Or more simply by looking at the number of new graduates in these disciplines (the flow) to the number of existing workers in these disciplines (the stock). In 2011, there was one new CS grad for every 100 workers in the field; ten years later that had doubled to nearly two new CS grads for every 100 workers.

Source: Author calculations from BLS Household Data (CPS) Table 11b 2011-2023 and U.S. Department of Education, National Center for Education Statistics, the Integrated Postsecondary Education Data System (IPEDS), 2011-2023 Completions

This all becomes even more clear in the chart below which shows the surge in hiring in the tech industry immediately after the pandemic. Not only were these firms hiring at a rapid rate, they hired much faster than was supported by output growth in these sectors. The gap between hiring and output really ballooned in 2022 and the first half of 2023. Then hiring flatlined, with employment and output growth matching each other only now.

Source: Burning Glass

Of course, there is always more to the story: some occupations within the computing field have expanded dramatically—information security analysts, for example, saw a 425% increase in employment from 2011 to 2023, while computer support specialists and software developers also grew steadily. Other occupations, such as computer programmers and web developers have plateaued/even declined since 2011.

Such dynamics reflect the volatile nature of the tech sector, which has experienced elevated layoffs recently (discussed above). Moreover, tasks that once required a full-time employee can now be completed by individuals/groups working short-term “gigs,” made possible through new tools such as AI (for instance, data handling and programing).

So, Is AI the Culprit?

One argument that has received a lot of attention places the blame for the college hiring slowdown on AI. It has all the elements of a great scare story. A shocking new technological change that promises to displace lots of workers; maybe the college unemployment numbers are the canary in an economy-wide coal mine!

We’re skeptical that AI is currently doing all the work that new college graduates once did. Recall that the unemployment gap between all workers and college grads has been declining since 2012, when generative AI was just a twinkle in Sam Altman’s eye and the stunning launch of ChatGPT was still a decade away.

And, some recent research on this point found no effect (yet) of AI on unemployment broadly. In fact, the authors found some evidence of just the opposite: while unemployment has gone up for all recent graduates, it went up the most for those whose jobs were least likely to be exposed to AI.

AI may well steal jobs from new college graduates in the future, but it hasn’t happened yet – at least not at scale. The authors also note that while we may not see a massive change in employment as a result of AI, we may well see such a massive change in ‘tasks’ – what employees actually do in their jobs. Obviously, this is a trend/development that is a must watch...

(To that end…hold the presses! The day after we finished writing this post, an important new study came out claiming just the opposite effect. Using high quality payroll data from ADP, Brynjolfsson, Chandar and Chen, show that there is an important distinction between jobs that AI automates and jobs AI augments. They find the former occupations (e.g. software developers, customer service reps) experienced a sharp drop in the hiring of 22-25 year olds beginning in October 2022. Prior to that date, hiring of young workers was proceeding at the same pace as hiring of older workers. Afterwards, hiring of young workers drops off sharply while older worker hiring continues. Note this is an unrefereed working paper days old. But it is provocative and an issue we will revisit in a future post!)

Some Implications

So, is the bottom falling out of the job market for new college graduates? Certainly not yet. Time for wholesale panic? No. A cause for concern/a deeper look? Yes, for sure. The soft job market for 2025 graduates and uncertain prospects going forward – whether transitory or structural - warrants action on the part of universities. Below are some ideas.

• We wrote on this point last fall: infusing relevance and professional skills into the student experience is likely the very best thing you can do for students going forward. We should start by asking: what’s the professional difference between a new college graduate and a 25-year-old? And, how do you get students to look more like a 25-year-old at age 21? Our new graduates are competing against people with experience and anything you can do to help them get more career-ready, closing the gap with those already working, can help them compete.

• Students need (and deserve) even more career coaching in this environment. While their classmates from just a year or two ago had their choice of jobs, it is tougher to land a job today. Helping students think through career options if their Plan A does not materialize is important – they need plans B, C, and maybe D!

• Universities and departments should take a deep look at their own placement data – do you have majors where students are struggling to find a job? What does this tell you about admissions/majors going forward – and adjustments you need to make?

• We have to watch the ‘hot major’ trends closely. Keeping your pedal on the accelerator in some areas like computer science and computer engineering may well be a recipe for a lot of students to be under- or unemployed given current trends in the market. (We know, we know – resisting the temptation to over admit in a hot major is much easier said than done!)

• There is enormous uncertainty in the economy right now – and assertions that industry has slowed hiring in response to that uncertainty make a lot of sense to us. Who knows when/if this uncertainty will clear? But, the deeper your relationships with industry/employers, the more likely you are to be in touch with what they see and what it means for their human talent needs – and your admissions, your career services, your curricula, and your co-curricular programs.

Upshot

Sensationalist headlines aside, the market for new graduates is softer than it has been over the past few years. This has happened before, and we assume it was called a crisis then too… That said, the uncertain prospects going forward strongly favor the institution that is in touch with what employers are thinking, aligning enrollments accordingly, counseling students appropriately, and doing all they can to help their students shorten the step into the work world.

Next in Finding Equilibrium

Next week, we take a close look at 5 myths about higher education – and do some debunking. Thanks for reading Finding Equilibrium – and if you have ideas for posts, let us know!

“Finding Equilibrium” is coauthored by Jay Akridge, Professor of Agricultural Economics, Trustee Chair in Teaching and Learning Excellence, and Provost Emeritus at Purdue University and David Hummels, Distinguished Professor of Economics and Dean Emeritus at the Daniels School of Business at Purdue.

Our co-author this week is Marley Heritier, a graduate student in agricultural economics at Purdue University.

To cut to the chase, who at the university is the ‘designated worrier’ – who is accountable for the employability of a student? If you can’t name a ‘designated worrier’ the answer is clear: no one.

Universities are complex beasts, comprised of a myriad of departments, colleges, centers, institutes, offices, and other units which exist because of discipline, tradition, budget, influence, opportunity,… Why they are so complex, and how that structure drives strategy (instead of the more preferred other way around) is something we will tackle in future posts, but here our focus is student employability.

Given this maze of academic departments, centralized student support services, co-curricular units, and placement offices, who is accountable for ensuring students are prepared for and secure the jobs/careers they want?

Career Services Offices are Not the Answer

No problem you say, we have a career services office – placement is their responsibility – problem solved! Not so fast…these centers play a crucial role in helping identify jobs available to students and working with students in a variety of ways to both choose wisely and put their best selves forward. However, many students don’t use these centers or use them sparingly: half the students in a recent IHE survey had never used career services or had visited them only once.

Source: Students Sound off on Career Centers, Inside Higher Ed.

Whether or not a student is employable is much bigger than the career services office. Employability – that combination of what students learn in the curriculum, their professional skills, and their personal attributes – is built over the entirety of a student’s campus experience, engaging many campus professionals.

Faculty worry about curriculum, ensuring high quality course content, and whether or not a student has mastered the material in their course. Academic advisors worry about whether or not a student has met the requirements to earn a degree/other credential. Academic affairs offices worry about providing experiential learning opportunities, tutoring services, writing support, and keeping students on a path to graduation. Student life offices support a plethora of professional development and leadership options for students, provide mental health services, and opportunities to engage in physical activities.

Multicultural programs offices build supportive communities across a variety of demographic lines. And, career offices do worry about making students aware of career opportunities they have identified and preparing students to interview for those opportunities. But, where do all of these offices, programs, and advisors come together for the support of any individual student?

Students: General Contractors for Their Own Employability

Given the siloed nature of all this support, the only true ‘designated worrier’ is the student themself. With no single office/person accountable for the student’s employability, students in essence are their own general contractors – trying to orchestrate all of the universities sub-contractors in a way that enables them land the job/career they want.

Think about a general contractor’s role in building a house or doing a major remodel: you know (or you will eventually discover!) you need a carpenter, a plumber, an electrician, a roofer, a painter, someone to pour your foundation, someone to install your flooring, someone to build your cabinets, a landscaper for plantings, … And, all of this requires some level of planning so an architect comes in handy!

The questions for a general contractor are endless: who can you trust to do the work, who do you hire, when do you hire them, which of the literally thousand options for every single choice (from cabinet hardware to paint) do you select? Then there is the business of seeking all the approvals, permits, and certifications required before you are ever allowed to move in…

Being a student at a large university is kind of like that – they are trying to orchestrate a very complex set of experiences that will help them land the job/career they seek. And, there are PLENTY of subcontractors to organize/sequence and PLENTY of approvals required to make that happen.

Subscribe now

Supporting Students as General Contractors

Some students, especially those of means and/or those with parents and/or siblings who went to college are at least partially equipped to play this general contractor role. Most other students, we would argue, are simply not and so do not take full advantage of the resources available to them. This makes them less employable than they could have been.

Don’t get us wrong: All of these programs, offices, and individuals mean well – they want to see students succeed and in all likelihood are passionate about student success. But students can be literally overwhelmed with ‘help’, don’t know where to start, and often get bounced all over campus in the process. The disconnected (and overwhelming) nature of ‘employability resources’ helps explain a fundamental issue: students are simply not aware of these resources.

Source: Driving Toward a Degree 2023, Tyton Partners.

A key problem here is organizational – too many resources coming at students from too many directions with not nearly enough guidance on how to put them to work. We have seen this especially true for underserved students who may be the beneficiary of so many programs that their (well-meaning) advisors are tripping over themselves to be helpful. This has always been difficult for students, but with the explosion of double majors, certificates, badges, and other cross-university programs, the challenges are even greater. Students get bombarded – and confused – by 10 flavors of the same ‘employability subcontractor’, each with different names, offered by a different unit on campus.

Some Ideas

If a goal is to produce and place more employable students, current organizational structures and coordination mechanisms deserve a careful look. Some ideas:

• Clarify your goals/strategy. Lots of universities organize around student success – typically defined as retention/completion. What does student employability mean at for your institution? How will you measure it? Suppose as a thought experiment you asked this question: how would we structure the student’s journey through the university if the principal goal was employability and progression through an outstanding career? What would we do differently? Responses to these questions will provide guidance as develop/revise your strategy for supporting student employability.

• Map student contact points. Building on the thought experiment above, take a typical student and identify all of the offices that are engaging with/could engage that student - beginning when you are recruiting the student. If you haven’t done this before, you will be amazed (or maybe shocked) at the number of offices, units, people who do or could engage with any particular student. What does this mapping exercise tell you about overlaps, blurred lines of authority, competing priorities, and just plain old confusion on the part of your students?

• Simplify where you can. Based on your mapping, are there obvious issues that can be addressed easily? Using common language, streamlining communications, combining duplicate efforts, improving coordination across units – anything that makes life easier for our student general contractor. While not a new idea, looking hard at co-locating ‘sub-contractors’ in a physical hub/one-stop shop can make life simpler for students. Likewise, a virtual hub/app that pulls access to such services together digitally can accomplish the same.

• Share information. Getting sub-contractors onto a common information platform can be highly impactful. Academic advisors know things about a student that leadership development professionals could use. Career counselors can do their jobs better the more they understand a student’s engagement in co-curricular activities. Our institution has been working on this issue for a while. Boiler Connect is an information system that gives those working with students access to all the information we can make available about that student. From giving student life professionals a much more complete picture of the student and their Purdue experience, to lowering student frustration levels by providing access to information they have already shared, this student-centered information system helps to deliver a far more personalized – and streamlined - student experience. Short of (and complementary too) such data systems, ‘councils’ or ‘boards’ that convene groups responsible for student employability can share information and coordinate efforts to deliver more streamlined approaches.

• Reorganize where needed. Are there offices/units/resources that should be centralized? Combined? Dropped? Less is more in many cases. Some reorganization may go a long way in clarifying responsibility/accountability and simplifying the lives of students. As one example, our University rolled a number of offices up under a Division of World Readiness including the Center for Career Opportunities, Office of Experiential Learning, Office of Undergraduate Research, Service Learning, and Pre-Professional Advising to drive much tighter alignment across these high impact learning practices and career services. The University of Tennessee-Knoxville Division of Student Success provides another illustrative example.

• Integrate employability competencies into the curricula. This deserves a separate post, but building partnerships between the academic departments and support units to ultimately embed career and professional competency development into the curricula and individual courses can play a major role in boosting employability. NACE provides many resources for a university working on this issue.

Structural Problems with University-Employer Relationships

Our focus here is the student side of the skills gap, but universities have the same challenge with respect to employer engagement. University development offices, research offices, careers offices, deans, faculty, … , all connect with employers. And, folks guard these connections fiercely (that incentive issue we discussed earlier). Given the fragmented nature of these relationships, it is unlikely that insights which might help the university turn out more employable students will find their way to someone who is focused on employability. Many of the ideas we provided for addressing the student general contractor issue could be used to streamline university-employer relationships.

Concluding Comments and What’s Next

Any efforts to remedy information, incentive, and structural issues won’t matter if there is no accountability. If students are employable who gets the credit? If they aren’t who takes the blame – and is responsible for making the needed changes to ensure they are? This of course takes leadership and it means making an outcome such as more employable students an institutional priority.

We hope you have enjoyed this series on the skills gap and what we can do to close it. We will likely return to the topic at some point, but there are plenty of other issues facing higher education that will get our attention this spring – starting next week. Thanks for reading Finding Equilibrium!

Subscribe now

Research Assistance Provided by Marley Heritier

“Finding Equilibrium” is coauthored by Jay Akridge, Professor of Agricultural Economics, Trustee Chair in Teaching and Learning Excellence, and Provost Emeritus at Purdue University and David Hummels, Distinguished Professor of Economics and Dean Emeritus at the Daniels School of Business at Purdue.

This week we tackle incentive problems that universities face when working with employers to narrow the skills gaps.

Let’s start here.  Students would be better prepared for long-term career success, and employers would experience smaller skills gaps in their preparation, if ....

Employers interacted more extensively with universities, sharing information about their needs and experiences with graduates, and sponsoring career relevant programming inside and outside the classroom.

Employer Incentives: ‘Trapped’ Students and Shared Information

For an employer, the ideal hire is a highly capable student who has exactly the knowledge, skills and abilities needed for a particular job… and is useless to any other employer.  Economists called this “match-specific human capital” – having capabilities that are only useful to your current employer.   Having match-specific human capital makes a worker productive in that particular firm but, lacking outside options, it reduces workers’ outward mobility and bargaining power.

Put another way, what is good for an employer is not necessarily what is good for a student. 

This incentive mismatch shows up in two distinct questions.  First, will firms invest in student programming or provide feedback about educational choices if that builds general human capital and makes students valuable to many firms?  Second, how closely should universities listen to employer complaints about students and their preparation?

As we have written previously, we think universities would benefit from deeper employer engagement to understand skills needed in the workforce that are not being provided by university programs.  And employers clearly have a strong incentive to hire graduates who will succeed in the jobs they are hired for and to identify the characteristics of both individuals and programs from which they graduate that correlate with that success.

But once this “secret sauce” is learned, what is an employer’s incentive to work with universities to reproduce this success on a broader scale?  After all, universities supply employers *and their competitors*.  Providing critical business insights that will benefit a competitor is not what successful businesses do.

This argument is doubly true when it comes to engaging in truly costly activity, i.e. not just sharing information but actively engaging the university in creating the human capital that an employer wants to hire.  Many employers are small and resource constrained and would need a big payoff to justify deeper engagement with universities.  But even for large employers, significant investments on a broad scale benefit many students who the employer will never hire.  

The related question is how much universities, or any part of the educational system, should listen to demands of specific employers for changes in curricula.  In our leadership roles we repeatedly encountered employers who would make the following series of claims/complaints.

1.       The university is not preparing students for the specific roles we want to hire, and so it becomes necessary to invest significantly in students’ training after the university.

2.       We are unable to hire students at the (low) wages we prefer to offer.

3.       Having hired students, and trained students, they are unwilling to remain in the position long enough to pay back their training and are constantly in search of new opportunity.

These specific complaints don’t strike us as the university’s problem, and we don’t believe that addressing them for these employers would be in the best interest of our students.

We are also wary of the recent (hard) push toward technical education, and especially moves to enable tracking high school students toward specific careers in manufacturing and skilled trades.  Yes, it would be nice for these employers (and for the students at least initially) to have a workforce of recent graduates who were skilled in the precise jobs the employers would like to hire today.  

But the low labor force participation rates and high disability rates among those who don’t finish college should make us cautious about these moves. 

And we should be under no illusions about what those employers will do if rapidly changing technology shifts needed skills or global market pressures push production overseas — employers will let go of those workers in a heartbeat or lower their wages to reflect the diminished market value of their skills.

All in all, this poses a real dilemma for universities.  We need to listen to employers, but not too much!  And employers need to engage with us, but not too much!  How to square this circle?

Share

Solutions: Engage Employers, but Not Too Much.

We think universities should:

1.       Make it easier for employers to engage and inform programming, but realize that advice given does not have to be followed.  This means we need to recognize the distinction between work requirements that are useful for very specific occupations or employers, and those that are generally valuable to many employers and across a student’s entire career arc.

2.       Seek other sources of information about needed skills. This could mean third-party intermediaries such as industry or trade associations or alumni who likely have more allegiance to the university and its current crop of students than to their particular employer. Alumni have the added benefit of seeing many employers and work demands along an entire career arc.

3.       Work with employers to expand micro-internships and the use of project-based courses where employers engage with small (3-5 student) teams to solve specific problems.  The associated work and deliverables provide context-specific practice of professional skills.  Students interested in a particular problem/employer can self-select into that engagement. Employers can invest more deeply in those self-selected students’ capabilities and hire them if that is desirable without worrying about investing more broadly. 

4.       Pair these employer engagements with an offer of skills certification opportunities for students.  Specific skill sets would not be required but the university could offer a menu for students looking to upgrade highly specific capabilities as they work with employers on project-based courses that require these skills. This would also signal to students that receiving a diploma is not the end of learning; it is merely the end of the beginning.

5.       Study these student-employer engagements occurring within the curriculum and co-curriculum to draw broader lessons for the kinds of capabilities that are broadly demanded in the work force and how their foundations can be better developed throughout students’ experience at the university.

Consider two examples of these more targeted engagements.  Many business schools offer case competitions or data-oriented competitions for teams of students that build all the most important professional skills.  In the best of these, a sponsoring company provides a business problem to analyze, proprietary data relevant to that problem, and a significant cash prize for winners.  Senior executives work with faculty to structure and judge the competition because it gives them a first hand look at how students think critically, solve problems, work in teams, and communicate their solutions.   At Purdue we greatly expanded these competitions and found them highly motivating both to students who upskilled, and executives who uncovered the talent they wanted to hire (sometimes on the spot!)  More often than not, students uncovered solutions the company had not previously explored.

Another example we really like is Purdue’s Data Mine, which has engaged hundreds of employers who are looking to upgrade their data science capabilities with thousands of students across all majors on the campus looking to develop their data science skills.  These employers work with faculty and students to design projects, provide large data sets and committed mentoring.  The student get project-oriented experience and motivation to tackle more challenging coursework that can assist with these projects.  The employers get deliverables to drive profitability, insights into faculty research capabilities, and a first look at future interns and employees.

For more ideas on impactful university-industry partnerships aimed at preparing students for the work world, check out the Business Higher Education Forum. This organization of business and university leaders offers a variety of resources on creating productive relationships between industry and academe aimed at developing tomorrow’s work force.

Better information and aligned incentives will help close the skills gap.  In our next posts, we will take a deeper look at challenges associated with the current generation of students, and then follow with thoughts on structural barriers to closing the skills gap.

Subscribe now

“Finding Equilibrium” is coauthored by Jay Akridge, Professor of Agricultural Economics, Trustee Chair in Teaching and Learning Excellence, and Provost Emeritus at Purdue University and David Hummels, Distinguished Professor of Economics and Dean Emeritus at the Daniels School of Business at Purdue. 

In our last post we addressed the incentive problems that universities and their faculty face as they work to improve programs and narrow the skills gap between what employers want and the graduates higher education produces.   This week we tackle incentive problems that universities face when working with students.

Let’s start here.  Students would be better prepared for long-term career success, and employers would experience smaller skills gaps in their preparation, if...

·       Students took challenging classes to push themselves and learn more, and engaged in non-required but still essential co-curricular opportunities to build essential professional and personal skills.

Student Incentives: Fun Today, Payoff Tomorrow?

At most comprehensive universities we provide students multiple pathways to a degree – not only multiple majors, but many variants of similar courses. Introductory calculus might be offered in four flavors: an elite honors section for future mathematicians, a rigorous course offered for engineers, something more applied for business majors, a cookbook version for students who never want to take math again.  Some of this variety reflects a desire to cater programming to individual student interests and specific degree paths.  Some of it reflects a wide range of student capabilities.  But some of it reflects a willingness to work hard in courses.  

And boy, as any faculty member knows, there is a lot of variation in the willingness to work hard!  

Every university has a set of exceptional students who make the institution proud in every way.  But, from what we and most faculty can see, far too many of our students do only what is needed to graduate, taking easy courses of study, cheating on assignments, and approaching coursework in the knowledge that “C’s get degrees.” Students may also believe that a degree from the university is the only signal that matters, and not the major, nor their grades, and certainly not their participation in non-required co-curricular activities.  What explains the poor work effort of these students?

One possibility is that there are generational differences in student attitudes and approach to work, exacerbated by the pandemic and its disruptions to norms.  We will tackle that possibility in an upcoming post, but let’s start with incentives that hold regardless of generations.

Working hard in courses and in the co-curriculum comes at an immediate cost.  That includes higher stress and cognitive load, the opportunity cost of working for money while in school (which many students need to pay bills), and the opportunity cost of missing out on the undeniably great social opportunities in college — football games, long weekends with a partner, and having too much fun at parties. But for all these costs, hard work brings a delayed benefit, with better learning and better career outcomes. 

In effect, there is a clash of incentives between the student today and the person that student will become decades later.  They are the same person, but the student deciding today whether to put in that extra effort is discounting the value to their future self. 

In this respect they are exactly like every other person who eats an extra brownie today despite their future-self wishing their waistline was slimmer or their A1C levels lower, or even their tomorrow-morning-self wishing they had stopped one cocktail earlier the night before. Students are bad at this because everyone is bad at taking their future self’s interest into account, a phenomenon economists call “hyperbolic discounting” – the “hyperbolic” part because people routinely discount the future so steeply its almost as if the future does not exist.  

That incentive problem is especially bad for younger people, and especially bad when the future is further off and uncertain. If students don’t understand how choices and effort made in college translate to improved career outcomes then effort feels like all cost and no benefit.

Share

Solutions: Take the Side of Students’ Future-selves.

The underlying behavior is not just a student problem (one of the authors just finished off the last of the Ben and Jerry’s carton.)  But a fix likely involves universities aligning themselves with the interests of the student’s future-self, not their current-self.  What would this mean?  Universities should

1.       Increase the salience of student choices by relentlessly connecting choices students make today to the tangible benefits they will receive a few years hence… and for the rest of their lives.  This could involve a concerted approach to advising, in which academic advisors and faculty repeatedly emphasize and encourage the right choices for students – starting before students arrive on campus.  Then colleges/departments could re-enforce the wisdom of these choices by creating opportunities for alumni (especially young alumni) to interact with students and tell their story. 

2.       Increase immediate benefits for working harder and trying more. This could include increased financial aid for high performers as well as heighted non-monetary recognition.  (e.g. Hummels used to offer a personalized recommendation letter from the Dean for case-competition winners.)  Or the benefit could involve creating direct connections to career outcomes, something like giving favored access to interviews with preferred employers.

3.       Create new sources of financial support for high-impact learning experiences such as study abroad and undergraduate research.  This is particularly important for students whose financial constraints require them to work during the school year and limit their engagement in the full university experience.

4.       Make individual courses tougher, not grade-inflating or otherwise providing easy pathways through courses, both to enhance learning and build professional skills such as persistence and meeting deadlines. 

5.       Reward, not punish, faculty for holding to a higher standard. A simple version of this would be discounting quantitative scores on student evaluations by the average GPA in a class, or at a minimum, reporting that GPA alongside the students’ approval ratings.

6.       Stop offering watered-down versions of majors that sound to an outsider suspiciously like the more rigorous versions of majors. 

7.       Get significantly more punitive about cheating.

These approaches would be costly to both students and the university in a variety of ways.  As an example, if these measures were not complemented with additional resources to support student success the university would likely experience increased time to degree and a declining graduation rate.  (A troubling recent study by one of our colleagues Kevin Mumford and coauthors claims that the sharp gains in graduation rates universities have enjoyed over the past decades is entirely due to grade inflation.) 

But in our view, it is past time to stop taking the side of our 19 year old customers and start taking the side of their 40 year old future selves.

Note: We won’t be posting next Friday, November 29, but will be back on Friday, December 6. Our topic: incentives and employers.

Subscribe now

“Finding Equilibrium” is coauthored by Jay Akridge, Professor of Agricultural Economics, Trustee Chair in Teaching and Learning Excellence, and Provost Emeritus at Purdue University and David Hummels, Distinguished Professor of Economics and Dean Emeritus at the Daniels School of Business at Purdue. 

In this post and the next we tackle the problem of incentives: what reasons do students, faculty and universities, and employers have to take action to close the gap?  How do misaligned incentives stand in the way of improved curriculum, course rigor, narrow versus broad preparation, and improved information sharing – and what can be done about this incentive problem?

Let’s start here.  Students would be better prepared for long-term career success, and employers would experience smaller skills gaps in their preparation, if the following changes were to occur.

·       Faculty designed courses and curriculum to provide students not only with rigorous disciplinary knowledge and skills (aligned with employer needs), but to build professional skills (communication, leadership and teamwork, …) by teaching in a manner that  provided repeated exposure to opportunities to practice those skills.

·       University leadership focused more on student outcomes and preparation for a career and less on athletics, rankings, or other measures of prestige.

·       Students took challenging classes to push themselves and learn more, and engaged in non-required, but still essential, co-curricular opportunities to build needed professional and personal skills.

·       Employers interacted more extensively with universities, sharing information about their needs and experiences with graduates, and sponsoring career preparation programming inside and outside the classroom.

We’ll tackle the first two in today’s post.

Faculty Incentives: Research, Rigor, and the Rest

Universities are multi-product firms, and we ask faculty to multi-task across four completely different activities that require completely different aptitudes. We ask them to produce:

·       research

·       students (teach courses, mentor, and advise)

·       service to the university (develop curriculum, administer HR functions of hiring, development, evaluation, and promotion/firing for our faculty, and sit on endless organizing committees)

·       service to their profession and/or to local firms, other stakeholders, and communities  (called  “extension” or “engagement” in land-grant and other universities with such an external focus).  

You might wonder… why would you ask faculty to do all these things?  And the answer is: because they have the subject matter expertise needed to do most of these tasks. How can an outside administrator or a staff member know what should be taught or what constitutes great research or who should be hired/promoted/fired in fields ranging from Accounting to Zoology?  They can’t, and so we rely on faculty.  Efforts to shift the task burden away from faculty toward less expensive staff members, particularly those tasks associated with running the university, gets criticized as “administrative bloat”.

How faculty allocate their time across these activities, and the quality of effort they put into each, depends a little on their intrinsic motivations but a lot on the incentives they experience within the university and within their profession.  Which of these tasks we incentivize varies considerably across universities, and increasingly across faculty types as hiring practices have evolved to distinguish the duties of more research-oriented faculty from those with more teaching-, engagement-, or administratively-intensive assignments.

At a research-oriented university (say, R1 or R2 in the Carnegie Classification), tenured and tenure-track faculty spend the majority of their working time on research, and their compensation is most closely tied to research output.  Some of this reflects internal evaluation mechanisms that reward research most heavily.  More of it reflects the fact that significant raises (and promotions) depend on externally visible research success.  

It is very hard to determine whether a scholar at another university is teaching well or offering great administrative service to the university.  It is relatively easy to see if they are generating large grants, placing books with top publishers or articles in top journals, and being cited in the scholarly literature.  And because it is easy to see these things, these achievements get rewarded on external (and therefore internal) markets.

A university could decide to unilaterally change this, down-weighting research excellence and up-weighting teaching or service to the university in order to get faculty to spend more time and effort on students and the institution.  A likely consequence is that the university would quickly lose their very best scholars.  A Nobel-prize winning economist we know once remarked, “My Dean doesn’t decide my salary.  He only decides which university will pay it to me.” In contrast, down-weighting teaching and service generally doesn’t cause a university to lose faculty who excel in these areas because these traits are largely invisible to the external market.

Share

Incentives to Teach Well

A starting point for any discussion on student preparation is the quality of classroom teaching.  We have 60 plus years of teaching experience, and having each just prepped a set of new courses for our return to faculty from administration we can say with some confidence: teaching well takes a lot of work!  

That is true of classical lecture and exam formats, but it is especially true when we make classes highly interactive and engaging.  And it is especially true when we use assessments (writing, open-ended problem solving, project-based assignments) that require more of students and build their professional skills.  Computer grade-able multiple-choice exams are a very low-cost assessment tool, particularly as class sizes grow to achieve cost-savings in the university.  Grading long essays for a 50-student course, particularly in a world where ChatGPT has fast become a ubiquitous part of the lazy student’s toolkit?  Give us a break.  

Introducing/maintaining rigor in courses is similarly disincentivized.  Why?  Suppose we want to take teaching really seriously.  What might we do?  

·       We would invest heavily in professional development for faculty, both creating institutional support for best practices and use of technology, and require faculty to participate. 

·       We would ask faculty to submit an advance copy of their syllabus and assessments to a committee of their peers to ensure that learning outcomes are clear and consistent with the broader curricula, and that assessments are fair but challenging. 

·       We would send other faculty to regularly observe the classroom to ensure that the style of teaching was engaging, and that students were motivated and participating.  

·       We would ask students to reflect qualitatively on the experience and how it could be better in addition to providing quantitative scoring of the course. 

·       Finally, we would use the outputs of this process to make decisions about faculty raises and promotions.

Subscribe now

Some universities do all the above.  In the universities we are aware of (in Denmark, for example), there is an enormous investment of time to ensure courses meet standards.  And the cost of that investment is significant:  faculty spend so much time creating, refining, and evaluating other courses that they only have time to teach perhaps one or two classes a year.   Every element of this more comprehensive approach to producing and evaluating teaching quality is expensive, with one exception: asking students to provide quantitative scoring of the course.  

And whaddayaknow, student evaluations are the most common way universities measure teaching quality and their quantitative summary measures the most important teaching input into raises and promotions.

What do we know about student evaluations?   Students don’t know what they don’t know, which includes knowing whether a course is preparing them for a great professional career. Students like fun classes and classes where they get a good grade. Students systematically downgrade instructors who don’t look like the authors of this article, with systematic differences in scores based on instructor demographics such as gender. 

Students’ future selves may be really grateful for tough courses that challenged them and made them think.  But their future selves aren’t filling out the end-of-term evaluations.  The 19-year-olds stewing over a C- on an exam are.

Share

Curriculum Not Courses

Student learning takes place across an entire curriculum, and a singular experience with one professor represents perhaps 1/40^th of the set of courses taken and an even smaller part of the overall college learning experience.  Curricular change requires coordinated action across dozens of courses and faculty, and the incentives to innovate there are weaker still.  Which is to say, faculty might be incentivized to improve their individual courses and especially bad teaching can have severe consequences including termination. 

But nobody is getting fired if the curriculum as a whole is an incoherent mess.

Related, there is an important distinction between faculty teaching well and students learning well that is relevant to this discussion.  Most universities provide some incentives for faculty to teach well.  But there are no incentives whatsoever connected to whether students actually learn material and can use it at some later date. 

The lack of incentives for student learning stems from several factors. First, the breadth and depth of material covered and assessed in an individual course means that the faculty teaching that course are usually the only ones who really know whether students are learning the material.  And for obvious reasons we don’t rely on faculty-reported data on student learning in their course to assess and reward/promote/terminate that same faculty member.

There are a few exceptions where student learning could be used to assess the quality of curriculum and instruction.  Some degree programs (accounting, medical fields, engineering) culminate in graduates taking comprehensive exams offered by bodies external to the university to assess the competence and fitness of the graduate to practice professionally. 

In addition, university-level accreditation bodies do ask universities to develop specific learning outcomes for their programs and assess whether students are meeting these learning outcomes.  As a practical matter, developing and using these learning outcomes (a very onerous process) to truly drive curriculum change doesn’t usually get far beyond the core curriculum.  Accreditation for specific disciplines can be more targeted, and loss of accreditation can have real consequences for programs.   

But we are not aware of instances in which such learning outcomes are used to evaluate individual faculty. 

Apart from the practical difficulties just raised, student learning is not specifically incentivized because learning happens outside individual classrooms.  It happens through the progressive accumulation of understanding over a sequence of well-organized courses and class-adjacent material in the co-curriculum.  It happens through a learning environment that includes meaningful peer-to-peer interactions.  

And to a surprising degree it occurs in courses that students think are not taught well at all!  (We recommend the book Range by David Epstein for more on this point.)  The fact that learning can arise specifically from creating “desirable difficulties” for students means that carefully designed randomized control trials have shown that there can be a negative relationship between our measures of teaching effectiveness and actual student learning.

All of this raises a profound question for universities.  Who exactly is responsible for assuring that students learn?   If everyone is responsible then as a practical matter no one is accountable.

Subscribe now

Solutions: Aligning Faculty and University Incentives on Teaching and Curricular Excellence

While the challenges outlined above are daunting we believe they could be addressed through the commitment of real resources.  We think universities should:

1.  Elevate teaching and learning success as a marker of the university’s prestige and reputation in the same way that research dollars, rankings, and athletics success are currently emphasized.  To use a line from one of our colleagues, universities should aspire to be recognized as “T1” (top tier teaching and learning) institutions, not merely “R1” institutions.  

2. Define what teaching and learning excellence means at the institution.  What is great teaching (and student learning) and how will it be assessed – in as objective and efficient fashion as is possible?  Use the definition to shape faculty professional development and assessment.  We are biased, but we like what our university has done in this regard.

3. Invest significant resources in institutional teaching and learning support to identify and promote best practices, broaden the use of advanced learning technologies, support experiential learning through investments in staff, and provide coaching and mentoring to developing faculty.

4.  Require participation of newer faculty in professional development courses and seminars aimed at improving course design and learning outcomes.  To ensure meaningful participation, we would reduce faculty teaching loads for the first year or two to ensure sufficient time to benefit from this learning process.

5. Increase rewards for and internal visibility of excellence in course design, innovation and delivery to place it on a par with research excellence.  As an example, in 2018 Purdue created ten 150^th Anniversary Professorships to be awarded on the basis of teaching excellence.   That was expensive – each professorship is funded by the university until the ten $1.5M endowments are in place – but it provided a profound signal about the university’s commitment to the teaching mission.  Of course, one could do this on a smaller scale to make it more affordable.  But the larger the award the stronger the signal.

6.  Elevate the importance of curricular and co-curricular design, innovation and delivery rather than only emphasizing individual teaching excellence. This could be done internally, with universities requiring units to regularly provide evidence of deep and thoughtful curricular review and/or awarding significant funds to winning departments for both innovative design and student outcomes.  It could also be done externally, with professional organizations and employers creating nationally competitive prizes  that highlight programs whose construction produces that kind of graduates employers aspire to hire.

7.  (Most controversial). Make faculty individually and collectively accountable and rewardable for enrollments and holistic student outcomes. 

At some level this is what RCM (Responsibility Centered Management) budget models do… they reward departments for generating student credit hours and/or majors.  We have seen this be most effective with new graduate offerings that incentivize faculty innovation by rewarding departments with a share of the tuition revenue generated from new programs. These funds can be used for research, to support PhD students, for summer salary, or other benefits faculty appreciate.  When faculty share in program success, programs become more successful.

But one could go much further with this.   Raise pools or one-time bonuses could be tied to student performance.  Faculty may strongly dislike this because while they are accustomed to being harshly evaluated and rewarded (or not) based on individual productivity, they are not at all accustomed to being held accountable for joint output.  But the reality in the work world is that bonuses flow to individual contributors when that contributor is productive AND the firm as a whole does well.  Why should faculty be different?

And, of course, we are seeing examples of faculty being held accountable for declining enrollments and student outcomes, but regrettably this occurs only when departments close and the faculty laid off. Presumably there are many opportunities to improve within programs not facing the death penalty!

Next week: students and employers and incentives.

Subscribe now

“Finding Equilibrium” is coauthored by Jay Akridge, Professor of Agricultural Economics, Trustee Chair in Teaching and Learning Excellence, and Provost Emeritus at Purdue University and David Hummels, Distinguished Professor of Economics and Dean Emeritus at the Daniels School of Business at Purdue. 

Collecting and Publicly Reporting Placement Data for All Students

Our first proposal is to require universities to collect and publicly report rich data on student placements.   Of course, universities typically do collect and report some placement data, but if you look at the details, you’ll realize that in most cases the statistics are based on a very small number of self-selected students.  And they are aggregated in a way that prevents much understanding of the variation in outcomes. 

Instead, universities should collect and report to state and federal data depositories the placement outcomes for every single student  with detailed information:  what degree they completed (using standardized instructional program codes);  are they heading to a graduate/professional school, still seeking, or have a job offer; and if they have an offer, what job they took (using standardized occupational codes), with what employer, and at what salary, along with information on debt incurred along the way.   This placement data (reported both separately for each campus in a university system, and aggregated over campuses) should be supplemented with data on non-completion (including debt incurred by non-completers).

To put teeth in such a requirement, we would incentivize participation by both students and institutions.  Completing this survey would be a graduation requirement for students - taxpayers subsidize the cost of education in both public and private institutions and deserve information on their return.  University compliance could be tied to state funding or to institutions’ eligibility to participate in student loan programs, with university leadership held directly – and legally – responsible for accurate collection and reporting of this data.

We would require that these data be made available anywhere a prospective student or employer or higher education funding body might look.  That glossy brochure highlighting the football team and the rec center might look a lot less inviting if it carried a warning label showing high loan defaults, low completion rates, or lousy starting salaries.  Funding bodies should know what their investments in universities are generating in terms of degree completion, jobs and earnings, and value for the economy.

In our view, this would yield one of two outcomes.  Either the data would support the view that universities and/or particular programs of study are adding meaningful value for students and for the firms who hire them - and help dispel a misleading narrative about the returns to higher education. 

Or, these detailed data would reveal significant weaknesses – particular schools, particular programs – where those returns are well below expectations, allowing students, administrators, and funders to make better choices about what should change and where money should flow.  Data of this sort could also help employers pinpoint hidden gems where they might more effectively recruit talent or understand why their sub-standard wage offers are being rejected.

Subscribe now

One subtlety here is that universities and programs differ markedly in the types of students they serve.   A university whose mission is to expand educational access by making admissions offers to many first-gen students, low-income students, and/or those with marginal SAT scores is very likely to see worse (unconditional) outcomes on completion, employment, and starting salaries.   This is the value of reporting data for every student, not just a self-selected sliver of the total:  data tools can easily filter results by the attributes of the incoming student population.  Funding formulas can be adjusted to reward the value-added of the university, that is, outcomes conditional on the characteristics of the students they admitted.

To see one example of what we are talking about, visit the Department of Education’s College Scorecard to see tools that provide information on earnings and student debt with information by degree program data.  The Post-Secondary Value Commission provides another tool building on the Department of Education data, supplemented with additional data sources such as the American Community Survey.  It provides additional nuance related to family income and race/ethnicity, and much better data visualization options for making comparisons.  

Share

However, a challenge with these tools is that they are based entirely on data from students participating in the federal student loan program.  For reasons we have discussed in an earlier post on student loans, this is an increasingly unrepresentative sample of students, as only 25% of students take out student loans, and they are more likely to come from low income families and/or attend more expensive universities.  Still, these are great tools!  Imagine what one could learn if we had systematic information like this on all students. 

Capturing Co-curricular Participation in Transcripts

Our second proposal elevates the role of co-curricular activities.  At most universities there is a clear distinction between “transcript-able” activities that are part of a required curriculum and everything else a student might do in the co-curriculum.  And as the saying goes, ‘what is measured is what is managed’.

It is clear to us that many of the professional skills students need for career success are developed outside the formal 120 credit hours required for a degree and are never formally captured.  Employers may see these listed in a resume but the university itself has no systematic way of knowing how students spent their time outside the classroom, nor any way to understand the role these “extra” activities played in placement success.  This makes it extremely hard to know whether these investments are sufficient, whether some or many students participate, and how participation correlates with student success.

Similarly, prospective students may be provided with a long-list of student clubs and activities as part of their college recruitment process.  But they have no way of knowing whether the university has made significant investments in co-curricular activities that are curriculum adjacent, no information that helps them identify which of these activities would best support their professional growth and hence generate the greatest ROI, and no formal way to demonstrate participation in them.

We will follow up in a future post discussing a few experiments (and associated challenges) with transcripting the co-curriculum.

Life-long Effects of College: Partnering with State Agencies

Our third proposal is that universities should partner with state agencies to link student-level transcript and related data to state administrative data from unemployment insurance or income records that provide a view into career progression and earnings throughout the professional life of a graduate.  Too much of what we know about graduate success for a particular university is based on first placement data.  To be sure, first placements matter, but we should know much more about whether we are setting our students up for lifelong success.

This would also help us to understand whether particular degree programs excel in first placements but not subsequent career progression, or vice versa.   It may be the case, for example, that liberal arts degrees do not equip students with specific practical skills that are in demand for entry level positions.  But that degree may provide capabilities that enable graduates to flourish in later career stages.  

An example of administrative linking in action can be found in Texas, with the extraordinary datasets maintained by the University of Texas at Dallas Education Research Center.  They begin with the universe of Texas high school students, capture their applications to (and acceptance/rejection) at 35 public universities in the state, and link it to workforce records to capture career outcomes - all with detailed demographic and school performance data.  This can be used by researchers who want to understand, e.g. the causal impact of getting into a better university, or changes in admissions policies that favor/disfavor certain groups of students.

If we could empower faculty scholars and institutional researchers with access to data that linked students’ curricular and co-curricular choices to their career outcomes, we could go much further in understanding correlations (or even causal linkages) to career success.  And we could more effectively experiment to understand how changes in curricula and/or student experiences yield different outcomes. 

Should state agencies prove uncooperative, a university could perhaps partner with LinkedIn to study career evolution or make a (much more) dedicated effort to systematically survey its alumni to learn more about their career progression and the strengths and shortcomings of their learning.  It is regrettable that alumni relationships are viewed much too narrowly and through a transactional lens:  inviting them to attend football games or basketball watch parties, or identifying their giving capacity and making a pitch for the annual giving campaign. 

Our alumni and their stories are perhaps our best source of information and inspiration and we know far too little about what in their college experience contributed to their career success, or lack thereof.

Share

Forward-looking Employment Forecasts and Employer Engagement

Our final proposal targets information that would be useful to university leadership in adjusting the mix of investment in, and graduates from, different departments.  And then adapting curricular and co-curricular experiences to reflect the particular skill sets needed.

Suppose a university president or provost wanted to better match the supply of students produced to the needs of the local economy.  How would they go about doing that?   They would need information on employers hiring plans, not just this year, but at least 5 years ahead.  They would need information on specific jobs the employers wanted to hire, and some information on the types of majors they would seek to hire from.

Most surveys of employers highlight gaps in particular capabilities like communication or problem-solving, and that is useful information.  But in our experience employers select for these capabilities within a narrow range of majors they might choose from.  And it is far more useful to a university to know employers want accountants who can problem-solve, and mechanical engineers who can communicate, if that is how employers plan to hire.  

There is some information like this available from the Bureau of Labor Statistics employment forecasts, and those forecasts have recently been supplemented to describe needed skills in some depth.  While these are useful (and all universities should at a minimum be studying this information), they have two weaknesses.  One, they are national in scope, and the majority of universities draw on a local student population who will, on average, mostly fill jobs in the state or nearby region.  Two, they are fundamentally statistical forecasting exercises operating at the level of the national economy.  Presumably local employers know much more about their future growth and hiring plans than a group of economists and statisticians in DC.

Share

Undoubtedly employers will push back against this suggestion and say:  we don’t know what the future will bring and we can’t tell you who we will want to hire.   Fair enough.  But if employers won’t tell universities who they might be hiring in the next five years, how in the world are universities supposed to produce graduates who will fill these hiring needs?

As a separate and equally important issue, university administrators, faculty, and staff need much more detailed information on the specific skills students will need to be successful in their jobs. We see this happening in two ways.

First, employers can actively participate as advisors to curriculum committees, faculty planning their courses, and staff supporting co-curricular activity. This requires some degree of sensitivity on both sides of that exchange.  Advice is not a vote, but advice consistently ignored will not be offered long.

Second, employers can help universities build the specific skills they want to hire by expanding internship and co-operative education programs.  Employers can also work closely with university departments to embed experiential learning opportunities within the curriculum as well as engaging deeply in co-curricular activities.  In our experience, these programs can be difficult to start and expensive to scale, but they can transform student capabilities and give employers a test-drive of future employees.

The University Role: Reactive or Proactive?

No university wants to have its financial resources tied to regulatory compliance, and many of our suggestions have more than a whiff of coercion about them.  That is in large part because value to students and to the university itself rests on comparison:  which universities and which degree programs are performing better and adding value and why exactly is that?   A reliance only on internal data which indicates seemingly satisfactory placement and completion rates makes it far too easy to sleep at night unconcerned with student outcomes.

So comparison matters and it is perhaps best achieved by some external pressure, either regulatory compliance or (shudder) by rankings. Still, even if faced with no such external pressure, we believe a university that took these proposals seriously would create a dramatically improved information environment and a starting point for making changes that could be distinguishing in the current environment. 

Of course, there is a risk that improvement would run into the wall of misaligned incentives and structural inflexibility, topics we tackle next.

“Finding Equilibrium” is coauthored by Jay Akridge, Professor of Agricultural Economics, Trustee Chair in Teaching and Learning Excellence, and Provost Emeritus at Purdue University and David Hummels, Distinguished Professor of Economics and Dean Emeritus at the Daniels School of Business at Purdue. 

Subscribe now

We begin a series of posts on this question with a focus on how fundamental gaps in information – what students, employers, and the university know about each other – lead to a mismatch between what employers want and what the university and students provide.  

Here is the sad state of affairs.   Employers don’t know enough about what universities offer. Universities don’t know enough about what employers want (either the mix of majors, or their content), or whether students’ pathways through the university prepared them for a successful career.  Students don’t know enough about how one university or one degree program differs from others, or enough about the experience of different careers and where they might flourish. This leads to a “build it and they will come” attitude on the part of universities, which largely works…until it doesn’t.

What Employers Don’t Know

For better or for worse, there is relatively little systematic engagement by industry in shaping university curriculum and the student experience.  Yes, some departmental and college curricular committees consult with industry experts, often asking alumni to serve in advisory roles.  But the heavy lifting of curricular construction, whether at the level of degree requirements, and especially the content of individual classes, is left to faculty.

This leads to significant information gaps.   First, industry professionals tasked with hiring college grads and supervising them in early assignments don’t know precisely what a degree program provides.  They likely know the broad contours of a degree, e.g. how the preparation of a Marketing major differs from a Finance major, or how Electrical Engineering and Computer Science graduates have broadly different preparation.  But they could not tell you the details of the required curriculum, and certainly not the content of individual courses. 

Perhaps more important, employers may struggle to map what is learned in a course or a curriculum to the associated skills students develop.  That is, faculty design a curriculum by asking:  what should students know?   Employers approach the end product and ask: what can students do?  Clearly, what you know affects what you can do, but how one maps to the other is not at all clear.

At some level this is understandable.  Employers recruit from many campuses and from many degree programs so knowing them all in detail would be a daunting task.  As a result, recruiters use place and degrees as a screening device, drawing perhaps on anecdata from previous hires and in many cases, alumni affinity. 

That is to say, employers are likely to know which programs produce strong graduates, but that understanding may not extend to the question of whether *this person* is highly capable, or whether *this degree program* bears strong responsibility through its curricular construction for that capability.   Further, judging by survey response data in which employers decry the capabilities of the students they chose to hire, this crude screening mechanism is not working well.   

What Faculty Don’t Know

Faculty don’t know whether their graduates are well suited to particular jobs for which they apply, or whether adaptations to curriculum would result in students better suited for particular jobs.  Indeed, most faculty would be hard pressed to answer the question:  what types of jobs do our graduates take when they finish? (Seriously, try this.  Go ask a typical faculty member who isn’t directly involved in undergraduate placement: “What are the top five destinations for your recent graduates?”)

Because faculty don’t know these things, they are unlikely to know much about job requirements, particularly when those requirements are evolving rapidly and differ significantly across firms and industries.  Their grasp on the mapping between course and curricular content (what students know), the resulting development of skills (what students can do), and the market demand for those skills, is weak to non-existent.  

Subscribe now

Faculty are also largely unaware of the central importance that professional skills play in students’ career outcomes, and as a result precious few courses provide intentional experiences to build these capabilities.  (One exception: requiring students to work, to write, to show up, and to meet deadlines may be one of the most important professional skills-building opportunities faculty can offer in regular courses. Getting back to that unflinching attitude and not putting up with a lot of nonsense would serve faculty and students well.)

Without detailed knowledge of how college experiences map into career outcomes, curricular and co-curricular design becomes an exercise in introspection, or a reflection of past practice, even if that past practice is increasingly ill-fit to current needs.   To be clear, in our experience it is not the case that faculty are unwilling to change.  It’s more a case of them saying:  ‘look, change is costly and change for its own sake can make things worse, especially if change is driven by some meddling administrators.  So, what is the evidence that a different approach would be better?’  

In our view this is a highly salient point: what evidence do we have that a different approach would yield better outcomes? In large measure, we don’t know, because we don’t collect data that would inform us on these points!

What Students Don’t Know

When students are choosing which colleges to attend and majors to pursue, they don’t yet know what their preferred jobs actually are, or their aptitudes for them. Nor do they know much about the experience of being in a particular degree program, or whether the choice of major X or Y or their choice of college A or B results in an increased likelihood of securing that preferred job. 

And that’s okay! A big part of college is to find out what you like, what you’re good at, and where you want to go.

But because they don’t know themselves or have much basis for evaluating outcomes, students tend to choose a college based on characteristics only weakly related to preparation for specific job outcomes, and majors and courses for much the same reason. Including, alas, because they are ‘easy’.

Worse, many students assume that curricular content is sufficient preparation for a rewarding career, and opt not to invest time in co-curricular activities that will build essential professional skills. We have had many conversations with students that go something like this… “If you thought it was really important you would require it.  The fact that these things are optional tells us they don’t matter.”  Which is, in some sense, completely backwards.  It’s doing the optional (co-curricular) things that really elevate students, and their capabilities, in the eyes of employers.  But somehow, students are not getting that message.

Is Everyone Really that Ignorant?

Given all this it’s sort of amazing that the ‘college to career’ pipeline works at all.  And it *does* work. Employers continue to hire college graduates, and those graduates continue to earn a significant premium in wages and employment relative to those who disdain a college experience.  

You might think we are overstating the ignorance of the central players in this drama. 

·       Students are increasingly jobs-focused, with their choices guided in part by placement rates and starting salaries.     

·       Even if individual faculty are not focused on placement data, academic leadership generally is, and they have hired a growing cadre of placement professionals to improve these outcomes.  We are also seeing significant shifts in the composition of majors, with small liberal arts colleges adding computer science, business, and engineering.

·       Industry professionals do learn whether individuals coming out of particular degree programs are more likely to be successful, and increasingly narrow their searches to institutions with a past history of success. And larger and more sophisticated employers are increasing the use of skills-based hiring to assess whether degree programs actually build the skills the employer needs.

Or you might think: as long as graduates secure employment, what’s the issue?   Perhaps placement success in the face of all this ignorance suggests we should be pleased with the broader value of a college education as opposed to focusing on difficult curricular change that might improve more specific preparation and student outcomes.

This sanguine view is misplaced. Nothing in the aggregate placement statistics speaks convincingly to the problem of a gap in skills preparation, one that might widen rapidly as the pace of technological change reorders job markets and fundamentally alters the returns to particular skills and capabilities.  Nor does it address the growing sense of dissatisfaction with preparation expressed by both industry and students. 

Far too many students begin college and incur expense and debt but do not complete a degree.  There is clearly an enormous gap in earnings across majors, even accounting for which institutions students attend.  The rapid rise in both skills certifications and enrollment in graduate programs that are more skills- and career-focused points to a broad dissatisfaction with the learning and career preparation that students received in their first four to five years in college.

All this suggests that curricular choices adapted to market needs really can generate better outcomes.  But the starting point for making better choices has got to be:  better information.   Next post: closing the information gap.

Subscribe now

“Finding Equilibrium” is coauthored by Jay Akridge, Professor of Agricultural Economics, Trustee Chair in Teaching and Learning Excellence, and Provost Emeritus at Purdue University and David Hummels, Distinguished Professor of Economics and Dean Emeritus at the Daniels School of Business at Purdue. 

We also believe universities that prepare more employable graduates will be living their mission of providing a broad, general education and preparing students to be engaged citizens, not compromising that mission nor turning into career and technical education programs (which are important for different reasons).  One thing is sure: universities that close the skills gap will earn favor with students, parents, employers, and elected officials.

Getting to Why

Solving any problem means getting to its root cause.  So, why does the skills gap exist?  That is a simple question without a simple answer.  First, there are multiple actors involved - students, universities (faculty, staff, and administration), and employers. Each of these groups is comprised of individuals and organizations with their own sets of motivations and constraints, some of them conflicting.  What is best for employers is not always best for students, and both students and employers struggle to understand the mission of universities extends beyond career preparation. On top of these factors, universities are large and complex places, which leads to incredibly complicated (and confusing!) interactions between students, the university, and employers.

Sorting out where in this complex system the breakdowns occur and where improvements can be made is not straightforward. Much of the criticism tends to focus on individual actors – students and universities especially.  We will outline some of these criticisms to provide context on where key issues reside.  That said, in this post and the next four posts to follow we will argue that a more important – and less obvious - set of factors are at play which impede alignment between student goals, university preparation, and employer demands.

Are Students To Blame?

Some argue the skills gap is the result of the current generation of students and their attitudes toward work (and life).  Employers say Gen Z brings an attitude of entitlement, they aren’t engaged or interested, they get offended easily, they don’t respond well to feedback, they lack work ethic, they lack motivation, they are hard to manage.   That sounds bad! 

But where do these attitudes and behaviors come from?  Some don’t blame the students themselves for their faults, they place the blame more broadly on ‘culture’, parents, and the pandemic.  And, even personal attributes can be affected, for better or worse, by the quality of the education students receive.

Source: Intelligent

Even if we focus only on students’ preparation for employment, we can find systemic problems that go far beyond student attitudes.  One study found that 50 percent of college graduates (two- and four-year) didn’t apply to entry-level jobs because they felt underqualified, with 21 percent indicating their college didn’t provide them with needed job-ready skills. While experiences such as internships are highly valued by employers, many students cannot access such experiences for financial or family reasons. In other cases, students may struggle to communicate the skills they actually have to employers. 

Something isn’t working.  Are universities and employers helping students to make informed choices about what to study, how to invest their time, and how to develop the skills employers want? 

Subscribe now

Are Universities to Blame?
Universities bear the brunt of the criticism for the skills gap, perhaps because students themselves don’t believe universities are adequately preparing them for the work world.   One study found students actually feel less prepared as seniors, relative to their freshman year, though this may be an example of “the older I get, the less I know” phenomenon.

Where in the university would one look to improve career readiness?  Perhaps the challenge lies with core curricula that fail to build foundational skills such as critical thinking and communication.  It might be that disciplinary courses/majors are not aligned with industry needs, either in the sense of providing the right mix of majors, or building the right combination of knowledge, skills, and abilities within a major.  Some critics have noted that universities do not emphasize building important employability skills throughout the curriculum and co-curricular activities.  Others suggest there is a lack of support and preparation for the career search process.

Criticisms aside, universities certainly make gestures, and sometimes enormous investments, toward addressing these concerns.  They organize curricula around ‘learning outcomes’ that focus on foundational skills such as critical thinking and communications. Colleges and departments have advisory boards to stay in touch with employers.

Universities also offer a myriad of opportunities in the classroom and in co-curricular activities to develop work world skills – undergraduate research, study abroad, experiential learning, roles in clubs and organizations, among others  They invest significant resources in career centers, interview workshops, employer days, cooperative study programs, and internship program support.

Still, something isn’t working.  Are these efforts sufficient and are they properly aligned with employer needs? Why aren’t students taking advantage of all the opportunities and career support campuses offer?

Are Employers to Blame?

Some critics place the blame for the skills gap on employers, arguing that employers hold down wages and push needed training back on universities – so the taxpayer can cover the cost of training instead of the firm.  They assert employers want to hire individuals with every needed skill already in hand instead of providing educational opportunities to remedy specific shortcomings.  Still others argue that there is simply not enough engagement by employers with higher education – that employers want to complain, but not invest in fixing the problem. 

For example, employers constantly state they want students with more work experience, but according to the Society for Human Resource Managers, 40 percent of employers provide no form of paid work-based training.   Many undergraduates must find paid work in the summer and/or during the school year to afford college expenses.  They don’t have the financial means to pursue an unpaid internship or other unpaid work experience. The same paper suggests employers do a poor job of onboarding students, which impacts the success of the recent graduate.

Source: SHRM, WGU

In employers’ defense, universities can be complex mazes – where is the entry point to make their needs known, or to learn about curricular offerings?  Small and mid-size firms in particular are resource constrained, which prevents deeper interactions with the university, makes more extensive in-house training difficult to arrange, and offers of higher pay or paid internships problematic.

Again, something isn’t working.  Are employers and universities doing enough to facilitate connections?  Who in the firm can best communicate what it takes to build a successful career to the university – or directly to students? 

So, Who is At Fault?

Placing the ‘blame’ on any specific actor – students, universities, employers – is misguided in our view because the problem is too complex for any part of the system to solve alone.  The good news is that everyone prefers a better prepared student, a more impactful university experience, a more satisfied employer.  The question is how to get there.

Drawing on our experience in administrative roles, we see four overarching issues that combine to create the current skills gap: information; incentives; structures; and generational differences.   We will address each issue in a separate post over the coming four weeks, including our thoughts on how to make improvements.

Where We’re Going

Information – Students don’t know enough about the likely outcomes of their choices, employers don’t know enough about the university curricula, and universities don’t know enough about what industry needs or how their graduates are faring.  How do we address these information issues?

Incentives – Incentive disconnects are evident across the discussion above – what is good for the student (a flexible skill set that prepares them for a variety of jobs) vs. what is good for the employer (deep training in a specific skill set).  Faculty have (some) incentives to teach their courses well but are rarely incentivized to focus on the overall curricula or student outcomes.  What incentives do students have to pursue ‘the above and beyond’ experiences that may help them develop their professional skills?  How do we find better alignment of incentives?

Structures – Who actually has responsibility for ensuring students are employable?  The ‘core curriculum’ at most universities is a misnomer, with little coordination between component parts flung all over campus, and little accountability for students’ mastery of essential critical thinking and communication skills.  Disciplinary departments are more likely to focus on capabilities necessary for graduate school than on skills needed to get a first job.  How do we address these structural issues?

Generational Differences – How is this generation different and what does that mean for how universities and employers engage with them?  Preparing students for success means acknowledging that some of the personal attributes in demand by employers can be cultivated through a college education - the question is how?

Over the next four posts, we will dig deeply into each of these issues and offer thoughts for addressing the challenges each presents.  We believe better and more timely information and the right incentive and organizational structures, which meet the current generation of students where they are, can begin to close the skills gap.  Next week: information.

Subscribe now

Research assistance provided by Marley Heritier

“Finding Equilibrium” is coauthored by Jay Akridge, Professor of Agricultural Economics, Trustee Chair in Teaching and Learning Excellence, and Provost Emeritus at Purdue University and David Hummels, Distinguished Professor of Economics and Dean Emeritus at the Daniels School of Business at Purdue. 

On the other side, students and faculty believe college grads are far more ‘career ready’ than employers do.  This difference in preparation/expectations has been termed the ‘skills gap’.  (There is a much broader question about whether higher education is providing enough of the right majors – sometimes termed the ‘skills mismatch’ - we will deal with this question in a later post.)

While we believe some of the criticisms of higher education are dramatically overblown, this disconnect between how college grads are prepared and what the work world wants and needs is real.  Higher ed can do better.  We will dig into this purported ‘skills gap’ in our next few posts – what is it, what evidence is there a skills gap exists, and importantly, why does it exist?

What is the ‘Skills Gap’?

Like underemployment, charges of a ‘skills gap’ or a ‘skills mismatch’ date back to at least the 70s and Richard Freeman’s book The Overeducated American.  Much attention is on the ‘professional skills gap’ (called ‘soft skills’ by many, a term not indicative of their importance): the difference between new entrants’ perception of their professional skills and the actual level of professional skills possessed by new entrants. A 2023 National Association of Colleges and Employers (NACE) survey is typical: 85% of college graduates surveyed rated themselves proficient with respect to ‘professionalism’ and 79% rated themselves proficient at ‘communications’, comparable figures for employers were 50% and 55%. 

We think the skills gap is broader than professional skills and like the idea of ‘employability’, defined as a student’s combination of hard skills, professional skills, and personal attributes.  Focusing on a student’s ‘employability’ lets us bring in the core curriculum, specific academic preparation in a major, professional skills, and personal attributes – all of which ultimately determine whether or not a graduate is ready for the work world and a successful career.  Unpacking the skills gap in this way provides insights into where the gaps in preparation exist and what can be done about them. 

Does a Skills Gap Exist?

While some question the existence of a skills gap, there is plenty of evidence of a disconnect between the preparation of college graduates and what employers want.  Many, many surveys focus on what’s important to employers, and how employers, students, faculty, and alumni feel about graduates’ level of preparation.

These surveys explore a wide variety of skills and personal attributes: some that can be developed in a classroom, some that can be developed through co-curricular experiences, some that can be developed through work experience, and some that are personal attributes of individuals.  Not all of these surveys deserve serious consideration: small samples, collecting opinions from individuals within firms who may or may not be familiar with new hires (CEOs, senior executives, etc.), etc.  But, those that do come to alarming conclusions.

Subscribe now

Professional Skills Gaps

The American Association of Colleges and Universities (AAC&U) has conducted a survey of employer perceptions for many years.  Key insights from this study are shown in the figure below – and the differences are gaping.  ‘Oral communications’, ‘adaptability and flexibility’, ‘critical thinking’, ‘complex problem-solving’, and ‘written communications’ show up at the top of the list – flagged as important by employers, and areas where students are not as prepared as they would like.  Utilizing technology, at the other extreme, is still relatively important, but students apparently have those skills in spades!

This study finds there are important age differences in attitudes toward these skill gaps, with employers under 40 placing more emphasis on ‘adaptability and flexibility’ and ‘complex problem-solving’, and employers 50 and older placing more emphasis on ‘oral and written communications’.  This finding provides insight as to where the emphasis on skills will be placed going forward.

An organization that brings together higher education and employers, NACE has developed a set of eight competencies that employers say are important to career success.  ‘Leadership’ shows up as the competency with the biggest preparation gap (32 percentage points), along with ‘professionalism’ (25 percentage points), and ‘communications’ (24 percentage points). 

A study conducted by APLU provides insight on how such professional skills build, from foundation through advanced, and explores opinions across employers, current students, faculty, and alumni.  Aligned with the NACE study, areas such as ‘understand role in the workplace and have realistic career expectations’, ‘recognize and deal constructively with conflict’, and ‘accept and apply critique and direction in the workplace’ show the most significant preparation gaps – and all three would be classified as ‘leadership’ or ‘professionalism’ by NACE. 

Common themes from these three studies are gaps in preparation in advanced skills, especially leadership and professionalism, and a foundational skill, communications.  We strongly believe college curricula and co-curricular activities can help students build these kinds of skills.

Gaps in Disciplinary Knowledge

General skills such as communications and leadership will always gravitate to the top of such surveys because these skills are important in virtually every job.  However, it is essential to note these kinds of surveys do not typically address a critical question: does the student’s major prepare them for the specific work requirements in that major - does a finance student know how to price an option; does an industrial technology student have appropriate CAD/CAM skills...?  Where available, research focused on industry-specific technical or hard skills shows similar preparation gaps.  Getting at career-specific preparation is a different – and equally important – issue.

Personal Attributes and the Skills Gap

The AACU study also does a nice job digging into the ‘mindset’ or ‘disposition’ of students that employers find important. These personal attributes also play a key role in the employability of the student and employer perceptions of preparation.  ‘Drive/work ethic’, ‘motivation and initiative’, ‘resilience and persistence’, and ‘self-awareness’ top the list.  A ‘sense of social justice’, ‘a desire to engage with the local community’, and a ‘worldview that allows people to understand the global implications of actions outside the United States’ drift to the bottom.  (We could see this set of personal attributes being very important for particular jobs.)

This set of mindset/disposition characteristics is personal to the individual student and perhaps harder to influence through the college experience – though it likely shows up in areas such as ‘leadership’ and ‘professionalism’.  That said, there would appear much more opportunity for the college that embraces this list as well and asks the question how these mindsets/dispositions can be cultivated through the college experience, relative to the college that assumes these traits are simply inherent with the student.

Closing Gaps

Some may be inclined to dismiss such feedback from employers: who were the survey respondents thinking about when they completed the questionnaire: the best new hire, the worst new hire, etc.; were respondents in a position to even evaluate students; we are placing students, what’s the problem; …? 

We don’t feel any college can ignore these kinds of findings.  Regardless of how well college graduates have fared in the past or how strong placement rates are currently, if colleges don’t adequately prepare their students for the work world, the work world will aggressively seek out alternatives to a college degree – and some of this is already happening. 

Colleges need to ask how they can close these gaps.  Do colleges, faculty, and students actually know what is important to employers?  How can these disconnects in expectations be addressed through the college’s curricula – and pedagogy?  How can they be addressed through student co-curricular experiences?  What about advising and counseling?

Importantly, how do colleges ensure there is more consistency with respect to career preparation across students?  We have personally worked with scores of students who were exceedingly well prepared when they graduated and went on to fabulous careers.  How does higher education ensure this is the norm and not the exception?  Finally, what is the role of the employer in helping communicate their needs and ensuring expectations are aligned?

Colleges must shorten the step students take from campus to the work world and we will dig more deeply into this issue with our next post.

Subscribe now

Research assistance provided by Marley Heritier

“Finding Equilibrium” is coauthored by Jay Akridge, Professor of Agricultural Economics, Trustee Chair in Teaching and Learning Excellence, and Provost Emeritus at Purdue University and David Hummels, Distinguished Professor of Economics and Dean Emeritus at the Daniels School of Business at Purdue.