I want to talk about the relations among AI, university teaching, and university budgeting. These links are not getting nearly enough combined attention.
Leonora Carrington, Crookhey Hall seen July 17 2026
College graduates are chief among the frontline victims of the AI story, joining non-college workers in a new phase of disposability (see Heck et al. 2026 for Brookings, with good Sankey graphics). The economic analyses of job market effects remain contradictory (Number 5 of my Director's Note for July, “Top Seven AI Trends for the Summer Break”), and yet data suggest that “young graduates face the grimmest job market in years.”
The human capital promise that learning leads (directly) to earning was never correct—earnings have always depended on a range of social and economic conditions (e.g. strength of unions), and salaries have always been set by firms and not by universities. But the wage benefit has been further destabilized by the AI industry’s claim that Large Language Models (LLMs) can surpass—and thus replace—all but the very best human cognition.
Economists see the relatively protected worker as the one possessed of “bundles of diverse skills” (Acemoglu 2025, my discussion). These allow the worker to perform “hard tasks” and not just “easy tasks” (predictable, thus more automatable). But what are these bundles, and how do you get them?
The basic answer is that the person who can do hard tasks can think independently of their tools, including a Large Language Model or other tech. The worker faces a problem, parts of which are poorly defined and about which information is incomplete and ambiguous. The worker has to have intellectual agency in relation to the definition of the problem and be able to make good choices about parameters and methods. The worker has to be able to combine heterogeneous elements and think across multiple epistemic frameworks (whose assumptions they need to understand). The worker must be able to admit and analyze error, be resilient with failure, repeat their effort with controlled changes of approach, take in criticism from others and also communicate in (complicated, sometimes adversarial) groups.
Universities discuss themselves almost entirely in terms of pecuniary benefits, usually measured by an individual graduate's personal salary. But their most important effects are non-pecuniary and social.
I set up a capabilities heuristic in The Great Mistake (before AI), and argued that every college should fund its achievement by every graduate. Last year, I modified it slightly for a presentation on AI to an interesting, sceptical group of Istanbulian engineers.
The brown squares mark the only steps that, in my opinion, can be enhanced with AI. All the others must be brain-only in their development, so that the resulting brain can use AI and not be used by it. The first step, "ability to study," appears to be endangered by the passive use of LLMs services. Same for "knowing what a research question is," which is trickier than it seems. And the heart of both learning and research, Steps 6 and 7, creating a well-formed research question and then forming a thesis or hypothesis about it, are precisely what students can bypass with LLMs (as this classic undergrad essay explained).
Universities are not set up to help everyone have many or most, much less all of these capabilities. Under decades of financial pressure, most stopped trying. As the latest and greatest ed-tech substitute for hands-on, face-to-face instruction, AI is causing cognitive damage whose variations are registered in new papers every day: yesterday's was AI's way of suppressing the ability to say "I don't know" in a way that enables thinking to continue. Academic advocates of AI increasingly insist on people actively using conceptual framings that stay independent of the models themselves (Andrew Piper).
There’s a large and growing literature of AI-related work about human capabilities and those that AI users absolutely need to have. Pretty much all of those I’ve read insist that workers need to control technology and not be controlled by it.
In his new book, The Reverse Centaur’s Guide to Life After AI, tech expert Cory Doctorow wants today’s “reverse centaurs,” like Amazon warehouse workers, whose powerful cyborg horse body dictates to their human brain, to be centaurs instead. In the process, he makes a particularly stark case for intellectual agency for all.
Automation isn’t necessarily the enemy of warehouse work: there’s nothing wrong with a forklift! The difference between automation that helps a warehouse worker and automation that torments that worker is whether the worker gets to choose where, when, and how to use that automation . . .
When you find yourself surrounded by people swearing that a given tool is worse than useless and others swearing that it has made their lives easier and better, you can bet that the former group is made up of reverse centaurs who’ve had AI imposed upon them, [and that] the latter group is all centaurs who’ve gotten to make up their own minds about where, when, and how to use AI tools. The solution to the paradox is to stop thinking about what the gadget does, and pay attention to who the gadget does it to and who the gadget does it for. (10-11)
Being a centaur requires massive labor of intellectual self-development so that one is basically smart enough to run the tech and not let the tech replace one's thinking. The university is a central agent in this creation of mind.
For intellectual power to exist at scale, graduates would have to be people who’ve not just had classes on “how AI is being applied within their fields” (p 13) so they can adapt to it. Graduates would have to be people with the capacity to put AI to uses that they themselves have conceived and designed. This means having intellectual agency over one’s “bundle of skills.”
This high standard flies in the face of widespread passive AI use. It also defies the history of capitalism’s replacement of labor (and labor’s judgment) with technology. This standard also runs afoul of the unbelievable capital investment being sunk into AI infrastructure almost entirely on the premise of imminent superintelligence that can replace untold millions of workers at a fraction of their cost. I am sure that this premise is wrong, but the enormous sunk costs are pushing the world’s most powerful capitalists into forcing it to be right. One way of forcing rightness is to run down the “median person”—and spread a technology that also insures their mediocrity by rotting their brains.
There isn’t a conspiracy, but there is a neo-eugenicist Valley disdain for the regular smart people of exactly the kind universities exist to enhance. (Check out Theo Baker’s Stanford saga, How to Rule the World.) The mind-boggling amount of capital consumed by the AI industry assumes the replacement (not the complementary enhancement) of mass quantities of human workers, which requires the industry to escalate their war on human capabilities. This means a war on the project of human development via formal education that runs from Plato to Kant, Fichte, and Schelling, on through Emerson, Douglass, Du Bois, Dewey, Jordan, Lorde, et al., continuing with current deep investigations in philosophy, software design, and cognitive science. Today's California Ideology substitutes AI for what it posits (erroneously) to be inadequate human intelligence.
This is an epic historical shift, from trying to make humanity more intelligent, by any means necessary including tech, to trying to make intelligence artificial, without really caring what happens to humanity.
Or to put it another way:
Doctorow’s apparently simplistic dualism that privileges the active intellectual use of technology is in fact supported by detailed scholarly analysis. I’ve especially benefitted from the thinking of Brian Cantwell Smith on reckoning vs. judgment (my review), Alan Blackwell on human agency with programming (my review), and Vivienne Ming on the complementary strengths of humans and machines, who notes, “exploring poorly structured possibility spaces remains a profoundly human talent” (23).
In spite of the variable and often fabulous uses to which we put AI (say, as every centaur’s Personal Assistant, Trend 2), it’s hard to see how big tech’s sunk capital will allow the AI industry to climb down from its categorical and unjustified minimization of human capabilities.
This in turn heightens the contradiction faced by colleges and universities. Their core function has been the creation of knowledge labor, with the default aim being “middle-class” stability and enjoyment. Around 1980, using the Bayh-Dole Act as a symbolic watershed, universities increased their visible service to knowledge capital. In addition to shifting science towards tech transfer and licensing returns (with important opposition from many science faculty members), university officials have in recent decades bought into every ed-tech capital substitute for knowledge labor. Now the AI industry is proposing the ultimate loyalty oath: universities must affirm the inevitability of replacing much—most—eventually all (depends on the day and the speaker) university-educated knowledge labor with AI.
Again, the point of the university, even economically, has been to graduate free workers rather than serfs to any company’s tech. That’s always been a key reason why people take the time, trouble, and expense to finish university.
For example, the University of California’s Undergraduate Experience Survey asks students why they selected their major. When they respond, four-fifths of students say “intellectual curiosity”; three-quarters say “prepare for a fulfilling career.” Only half pick “leads to a high-paying job" (Table 1, p 6). The pecuniary benefit is baked into a college degree; college is an exhausting, expensive way to get a decent salary if you don’t also really want to acquire intellectual strength so you can have the challenge and pleasure of grasping reality and affecting the world.
How are universities going to cultivate “human judgment, creativity, and subject matter expertise,” as UCOP put it this month’s report, “The Economic Impact of a UC Degree”? Many people will say that you just need decent skills with using AI so you can set up a rig that works for you. I love how many smart and dedicated people are sharing tips on excellent chatbot workflow transformation all over the internet. But all of these setups can suck out intelligence rather than extend it.
All users need to have mental strength going into AI use, and continue to develop that strength. And that requires tremendous (and ongoing) offline, brain-only work. This work is the point of colleges and universities. They need to offer the regularity of back-and-forth cultivation that only elite universities have even tried. The need to furnish tutorials and small-group learning for the masses, for the first time. This will start with much smaller typical class sizes than publics have right now. You can no longer have classes with 140 students that you call “discussions.” Even 40 students is way too many for deep individual development. Large lectures can no longer lack discussion sections, and departments will need to fund face-to-face interactions on which teaching and assessment is based. How will, a university like UC Irvine pay for meaningful assessments when they should no longer have TAs grade 300 papers written on laptops with multiple AI subscriptions? Universities now need to offer group work that’s engaged and supervised by advanced faculty, on the model of crits in fine art--with help via reassigning middle managers to the educational core. In contrast, most students now appear to be using LLMs to complete take-home exams and problem-sets, making these useless: for just one example, see “How a Blind Professor Saw Through His Students’ Cheating” at Brown University. Universities will need to develop oral tutorials and exams, with tech assists monitored and complemented by new intensities of face-to-face contact between students and instructors.
Here we run into a big problem. Any meaningful combination of these upgrades will bankrupt the current business model. The model has been draining funds from instruction for decades, through vendor outsourcing, adjunctification, administrative growth, capital projects, and the many varieties of “mission creep.” Universities with money to spend, spend it on administration. At MIT, “faculty grew by only 9.2% between 1985 and 2023, while administrative staff grew by 189%.” This is a decades-old problem that has led to diluted instruction, against the talents and the wishes of the contingent faculty who teach the majority of the nation’s college courses by rationing their per-student time to survive. Universities have never addressed the administrative bloat, the reduced per-student instructional resources, or the steady one-way adjunctification. Only union bargaining has slowed the drift.
By the 2010s, the U.S. university system had locked in reduced investment in students that also varied by racial group (Black, Indigenous, and Latino students go to poorer schools than Asians and whites): they had set instructional funding at less than one-third of overall expenditures (see Nate Johnson on both points, comparing Figure 1 to Figure 3). Teaching at public colleges is highly dependent on state funding and tuition: at UC Irvine, for example, 70% of “core funding” comes from these sources this past year (and 81% the year before). Nationally, state revenues for public colleges are still 2 dollars short of their (inflation-adjusted) level of 1999, having spent most of that period below the level at the turn of the century. Even with a 50% increase in (inflation-adjusted) tuition, total revenues have only increased 14% over 25 years, with likely all of that increase--and more--going to non-instructional functions.
In short, the typical public college or university lacks the money to improve instruction to the point that the vast majority of its graduates can run AI and not be run by it. Non-wealthy privates are in the same position.
The UC report I linked above, on the "economic impact" of a degree, ducks the question by saying the current system is working fine because look at the graduate salaries. UC grads do have better salaries than non-graduates and the graduates of some public colleges. It is indeed still better to have a B.A. degree than not. But a UC graduate's earnings tells you little about their actual learning on a UC campus: it tells you about labor markets, and in this case labor markets for people who left college 5-10 years ago, in the pre-AI era. The cognitive crisis must be faced directly, and so must the need for new (public) investment--a new political economy-- to support real solutions to it. I'll discuss this issue in Part 2.














