Universities in the Shadow of AI

To recap: in the Instrumental University I considered what the government wanted a university to be. Then I looked at universities as load-bearing institutions of democracy. Big hint: the two have a nodding acquaintance, but do not align.

The second of the two touched on some of the forces that were seeing the university unbundle itself—and so becoming less than the sum of its functions. The last factor forcing a further un/re-bundling is technology, in the form of AI. AI is having a demonstrable effect, undermining credentialling, skilling and knowledge production and transfer—thus attacking directly the university’s economic rationale.

The first order effects are the things we can already see.

The collapse of assessment integrity. The essay, take-home problem-set—the consistent process of verifying-that-learning-occurred is broken. Universities have deployed their first responses, aided by coping with COVID: detection software, a return to proctored exams, and viva-style assessment.

However, each of these increases cost per student, much more than the assignment-marking they replace—directly attacking the mass model’s unit economics. In other words, the cheapest thing to fix the integrity problem is the thing that breaks the business model.

Content-delivery value goes to zero. The lecture was already dying, through recorded lectures and streaming, again hastened by COVID. AI finishes it: a patient, infinitely available, personalised tutor beats a recorded lecture on every dimension except credentialed authority.1

Students are already routing around the university’s teaching for the actual learning

Academic labour restructures. AI does the literature review, first-draft writing, grant-boilerplate, marking, and codified work. That may sound like a productivity gain, but it hits the junior rungs the hardest—the research assistants, the postdocs looking to pay their way through, the tutors and sessional markers, the early-career researcher doing the codifiable grunt work.

But these are the rungs on the apprenticeship ladder, the tasks through which vertical enculturation happens—the learning of the craft. So while using AI to substitute for the junior labour sounds efficient, it quietly severs the reproduction mechanism; the learning and refining of repetitive motion and tight feedback loops. That’s a first-order efficiency creating a second-order sterility, and nobody’s costing it because the pipeline damage shows up in 15 years or so—not this budget or election cycle.

Research production accelerates and degrades simultaneously. Similarly, AI speeds hypothesis generation, drafting, and coding. And it floods the literature with plausible, fast, sometimes-confabulated output.

So: more papers, faster. But the problem of provenance becomes acute—the signal-to-noise ratio of the knowledge commons begins to swing down, which raises the value of trusted authentication even as it lowers the value of raw production. The university that keeps optimising for production (papers, grants, output metrics) is optimising for the thing going worthless while neglecting the value of things going scarce.

Share

Second-order effects are where systems die, because they’re the adaptations to the first-order shock that dissolve the thing you were trying to save.

The cross-subsidy load-path fails. This is a problem inherent in the Australian model, and potentially its killer.

Currently, international-student teaching revenue cross-subsidises research and domestic teaching. What happens if the credential’s signal degrades because verification is broken? The international-student proposition—pay premium fees for a trusted Australian credential—weakens at exactly the point of sale.

What happens if AI delivers the skilling an international student could get anywhere, cheaply, in their home country? The ‘come to Australia for the education’ value proposition thins to ‘come to Australia for the migration pathway’, which in reality was often the real product, but now it’s nakedly the real product, stripped of educational cover, which makes it politically indefensible and regulatorily exposed.

So AI doesn’t attack the cross-subsidy directly; it strips the educational justification from the revenue source that funds everything else, leaving a migration-pathway business wearing a university’s clothes, vulnerable to the first that decides to see it plainly. And the migration politics is already live (the Australian caps, the visa tightening, the anti-immigration sentiment being fanned by One Nation).

AI strips away the institutional veneer. Once that facade collapses, the cross-subsidies supporting both research and domestic teaching instantly evaporate.

Cost structure and value proposition invert against each other. As mentioned, the first-order fixes (in-person verification, viva assessment, dense enculturation) raise cost per student. The first-order value erosion (content delivery worthless, skilling commoditised) lowers what students will pay for the mass product.

So the sector is being ‘scissored’ by rising costs on the defensible functions and falling revenue on the mass functions. There are two choices ways forward: charge a lot for the dense, verified, in-person thing to those who’ll pay; let the mass function go to cheap AI-delivery.

Which means the second-order effect of AI is to make the current mass education, affordable, social-mobility university financially impossible—not because anyone chose to abandon it, but because its cost floor rose and its price ceiling fell in the same motion. The post-war/post-Dawkins settlement of the Australian university system gets priced out of existence.

Institutional differentiation collapses toward reputation. When AI commoditises content and skilling, the only remaining differentiator is trust—and trust is a reputational stock accumulated with consistency over decades, especially at an institutional level.

That means the value concentrates in the institutions with pre-existing reputational capital (the Go8, globally the elite brands, with one or two others, such as RMIT), because in a trust economy, the incumbent trust-holders win by default.

AI, sold as a democratising, access-expanding technology, produces reputational oligopoly in higher education—the rich-get-richer dynamic of trust markets. AI mechanically concentrates the surviving value into the already-trusted.3

Then the state’s calculus changes. If AI can deliver mass skilling cheaply, the state’s instrumental reason to fund mass universities (workforce production) weakens—AI does it cheaper. So the utilitarian funder looks at the university and asks ‘what am I paying this for, when the AI tutor skills the workforce at a fraction of the cost?’ The answer is ‘you are paying for the functions that AI cannot do—provenance, verification, independent expertise, custody of contested knowledge’—that is, the essentially democratic functions of universities.

But those are the functions the utilitarian/instrumental approach to universities values least and distrusts most. The workforce argument that sustained public funding since 1945 evaporates, and what remains is an argument about democratic infrastructure that has, as established earlier, no organised constituency. AI removes the utilitarian case for the university at the exact moment the democratic case is the only one left—and the democratic case is the weak one politically.4

Such paths are not linear. For example, those academics and researchers who most embody the independent-expertise function—the ones with international options—are typically the first to leave a garrisoning, utilitarian, low-trust system, because their value is portable and their tolerance for instrumentalisation is lowest. That can lead to a loss of critical for core knowledge and expertise, often invisible until it is needed.

Changes happens slowly, until it happens fast. And there is still time to adapt. In the next piece I’ll be looking at the prospects for change and adaptation, because the Australian system—and its politics—has some structural peculiarities and path dependencies that will need to be addressed directly.