This title may seem odd as the British job service Indeed reports that entry-level job postings are down 7% this summer, to their lowest level since the 2020 pandemic. 

It may seem odd coming from me, since I often argue that universities made a fateful conceptual and strategic error when they put themselves on the hook for graduate jobs and salaries, which are controlled by business. The value of a university degree is mainly non-pecuniary and public. Universities are really about intellectual development, new ideas, and social knowledge.

But universities must also do no harm to employability. Tragically, when they adapt their already narrow job focus to AI—instead of piling their money into full student development—they do that harm.

In our AI purgatory, the issue for post-graduate jobs is whether the worker has been taught (in large part at university) to follow AI or to run it. Cory Doctorow insists that our debate about whether AI is good or bad is really between people who control AI as a tool (good) and people who are controlled by it (bad). Amazon warehouse workers are Exhibit A for the latter, where AI-enhanced control systems deliver famously humiliating conditions and the worst injury rate in the business. This dichotomy—controlling AI for your benefit vs being controlled by it—does explain most of the divided landscape on AI’s overall effects.

Are universities really teaching the difficult intellectual capabilities that will allow graduates to control AI, both to automate tasks within their future jobs and also to do hard tasks without it? Or are universities adopting industry BS about AI having general intelligence superior to that of their own graduates, going “AI Forward,” and teaching dependence on AI? 

Who would do the second, you might wonder, until you read university statements that encourage widespread and uncritical use of AI as a future job skill. Or read tech influencers like Andrew Ng advising young developers to use AI in 100% of their coding, rather than learning to pick and choose when and how to use it. Entry-level employment has declined in computer programming, and full conformity to the tools is a recipe for further worker replacement.

You might wonder who would teach AI dependence until you notice the extent to which universities are treating critical disciplines as disposable. They routinize and simplify learning by laying off academic staff, at Exeter and Goldsmiths and large numbers of other UK universities. They imply that writing is something LLMs do when they close a prominent writing programme and sack its director, an internationally respected theorist of AI and writing at Harvard.

The silver lining in the university’s AI cloud is that the power of the models will force universities to choose between two contradictory theories of what education is for.

The power of the models will require active intervention if we want to fulfil dissenting visions of the future of work. One example is Branko Milanovic’s Marxian observation that capital-intensive sectors (AI) logically must sustain labour-intensive sectors (nursing, creative writing) to keep their rate of profit from falling to zero. This is broadly correct in the abstract, yet has to be made true by various public actions. These include university practices, which must help form the kinds of people who can choose and control their use of technology in the workplace.

Since we’re in the age of apocalyptic statements, here’s mine: universities are generally choosing the wrong theory. If they continue they will slowly die, while also helping the AI industry deal with the tendency for its rate of profit to fall to (below) zero by installing a lasting technofeudalism.

One theory of university education is that education teaches students to internalise what is known and apply it faithfully. The second theory is that students must also learn to think independently of the material and its system or the practice isn’t really education.

The first theory is legitimated by the employability agenda and is rooted in human capital theory: learning means acquiring the specific knowledge that increases your workplace productivity, which a free market duly registers with a higher wage. The second theory is that education slowly cultivates intellectual autonomy. It is associated with forming democratic subjectivity (John Dewey, W.E.B. Dubois) and everyone’s powers of creativity.

The first theory is measured by wage effects broken down by field of study (Britain’s infamous Longitudinal Education Outcomes metric, the University of California’s dashboards and reports on “the economic impact of a UC degree,” etc.). The second resides mainly out of public view, in various corners of philosophy, cognitive science, computer design, linguistics, media studies, liberal arts theory. It also lives in students’ stated preferences (Table 1, which shows that more students chose their field on the basis of “intellectual curiosity” and career fulfilment because it “leads to a high paying job”).  

The first theory directs education to serve the economy. The second expects education to expand the cognitive and affective capabilities that improve life in society. Universities use the first theory to bind degrees to employability, which dis-embeds learning and thinking from culture and society (the “civic university” tries to bolt community service back on later, but these public effects are not seen to emerge organically from the university’s academic activities). University officials tend to shun the second theory in which education forms intellectual agency for collective use (as being nebulous, unquantifiable, controversial, unbusinesslike).

The first theory tolerates the steady withdrawal of funds from instruction on the grounds that the desired outcome is straightforward (job readiness) and the curriculum can be narrow and direct. The second theory costs serious money for more instructors, more interaction, more contact, more trial and error, more feedback, more module sequencing, more individual study, more structured group work. 

Ironically, the first theory—learning for economic service—radically depends on the (often disavowed) presence of the second—learning for cognitive and interpretative agency. Intellectual autonomy is the driver of the capitalist valorisation process as well as a source of criticism of it. Mainstream theories of the “information economy,” “knowledge work” and the “creative class” all acknowledge this point, which has been extensively developed in Marxian and non-Marxian models of cognitive capitalism. Universities ask students to focus obsessively on courses that increase their employability while silently assuming that this will also produce better capabilities. Graduates on the job, though typically still in their early 20s and starting out in life and work, are indeed expected to exercise judgment, decide among various work plans, continuously redirect effort in response to new information, and work creativity with others as well as on their own. The graduate, in reality, will not be asked only to adapt to systems but also to evaluate, change, and (re)build systems as the environment changes. Nobody wants to hire someone who absorbed the coursework and applies the rules, no matter how well they did in school.

What discourses are strong enough to oppose this? The smart-AI /servant-worker view, building on the old theory of worker as cog in a machine, has always been contradicted in the humanities, arts, much of the social sciences, and also by science educators and nearly all of the working scientists and engineers that I have known. I can’t recall meeting a single student or academic who espoused routinised higher education for subservient efficiency at a future job. Meanwhile, university officials default to Theory 1 in which university education is training in adaptation, not to say conformity to whatever business says it needs for the good of the economy.

Yet such officials also listen to mainstream economists. Interestingly, many such economists who study knowledge work and AI are reinforcing the “Theory 2” arguments for education as “growing people.”  

I’m going to extrapolate some rules of thumb for working under AI from a set of mainstream economists, including Daron Acemoglu (whom I previously discussed in this space) and a new book on AI labour, Messy Jobs: The Work That AI Cannot Reach, by Luis Garicano, Jin Li, and Yanhui Wu. I invoke these prominent economists because they are the kind that university officials can hear.

1. Jobs aren’t undifferentiated units to be replaced or not. They are “bundles of tasks,” and most jobs combine easy and hard tasks. I would say all jobs are like this, including “unskilled” jobs, but it’s clearly true of all jobs that require a university degree.

2. “Hard tasks” are those that aren’t sufficiently predefined to allow reliable automation. Their goal isn’t fixed and specified, the means for achieving it aren’t spelled out in advance, or both. Hard tasks often take a long time to learn (years). We all have experience with a related feature: these tasks, like writing a monthly Note, are sufficiently different each time so that they don’t really get easier to do. Continuous learning, interpreting, and changing are required to do them well.

I’d suggest again that most jobs that add a lot of value are like this—not routinisable—and the key to doing them is one’s cultivated general capabilities rather than settled procedures. Garicano, Li and Wu stress the potential for mediocrity, work slop, verification problems, and trust collapse when hard tasks and jobs are automated.

3. Modern work is collective, so a job’s bundle of tasks will always involve interpersonal, situational, and social knowledge. The humanities and their “soft skills” are having a moment for this reason. But social capabilities have always been important in capitalist labour: one orthodox measure is historical wage data, which suggest that “social skills were not a substitute for analytical ability, but they were a multiplier of it” (Garicano, Li & Wu 57).

4. Organisations and their workers have to decide, actively and empirically, when AI can and can’t automate without important losses of quality, and when AI can and can’t be used to coordinate tasks. Contra Hayek and all his bastards, markets don’t organise work correctly: organisations do.

5. Contra the AI evangelists, AI is best for “simple” tasks and people are best for “delicate” (complex and/or interpersonal) tasks. This is not because of a bright line between humans and machines in which humans have species defining creative uniqueness (see the impressive results of this starting point in Leif Weatherby’s Language Machines) but because of the properties of the hard job bundle (ambiguity, incompleteness, inter-personality, situatedness, reception of real-time human feedback, accountability, higher costs and/or quality drops from separating tasks out for automation). Messy Jobs has good examples of firms that use LLMs extensively even as “human agents” continue to play “an indispensable role” (e.g., 53).

6. There is no general market logic in which AI technology lowers labour costs such that we get a job apocalypse. Our corporate masters may well induce a job apocalypse, but it won’t be the models that made them do it. This is true for various reasons, including the high marginal costs of LLMs and also the drivers of the falling rate of profit of capital-intensive industries (here Milanovic is joined by the Bank of International Settlements). This is true of “cognitive” labour as well as hands-on trades.

7. What is happening and will happen is AI-related job change. As Weatherby puts it, “if you touch a computer for some material reason in your job, your job is going to change. If you use a computer for something you consider important during your day, your day is already changing.” Garicano, Li, and Wu take a step beyond Acemoglu in describing jobs that will change but continue as “strong bundles,” and jobs to be farmed out to AI as “weak bundles.” The difference is that in the former case, breaking up the job destroys value. How you know when this will happen is complicated—it can’t be answered by invoking the rapid evolution of LLMs.   

The economics terminology won’t appeal to non-economists (you are perhaps already tired of the words “bundle” and “tasks”). But the basic idea is that strong bundles (that workers must keep) have variable inputs, interconnected effects in which one task affects many others, and output qualities that can’t be easily metered. These and other features mean that such jobs require ongoing, complex worker judgment in assessing and redirecting them.

8. Workers need to control the unbundling and re-bundling of their jobs. This is the main alternative to the mass sackings that corporations have indulged in since the late 1970s. It of course requires organisation—unions, guilds, movements of the cognotariat. But it also requires advanced consciousness, “cognitive sophistication” across a range of technical, cultural, social, and organisational issues.

Either universities will take on the job of developing these forms of intellectual autonomy for all graduates, or workers will look elsewhere for help. In the latter case, university decline will continue. But that fate is completely unnecessary.