AI and Higher Education

A university lecture hall with a laptop open, showing an AI interface

Section 1 — From School of Thought to School of Profession

The idea of a place dedicated to higher learning is not a Western invention, and it is far older than the university as the term is understood today. Plato's Academy in Athens (387 BCE) was one of several such foundations across the ancient world — schools of thought before they were schools of anything else, places to study philosophy, rhetoric, and the natural world for its own sake. China's Jixia Academy, founded within a few decades of it in the fourth century BCE, arose with no plausible contact between the two civilisations: a genuinely separate invention of the same idea (Dukes Plus, 2025; List of Oldest Higher-Learning Institutions, Wikipedia, 2025). The others that follow sit on a different footing. Nalanda, in India, grew from the fifth century CE into a residential centre drawing students from across Asia to study logic, medicine, and Buddhist philosophy alongside the Vedas, by which point Hellenistic and Buddhist thought had already been fusing across Gandhara for several centuries (The Chopras, 2023; Ancient Universities of India, Wikipedia, 2025). The Jewish academies of Sura and Pumbedita, and later the great yeshivot, sustained centuries of legal and textual scholarship through the Talmudic tradition, in a Mesopotamia that had absorbed Hellenistic administrative habits under earlier Seleucid rule (Sura Academy and Pumbedita Academy, Wikipedia, 2025). Al-Qarawiyyin in Fez (859 CE) and Al-Azhar in Cairo (970 CE) show the clearest such line of all, founded after the Abbasid translation movement had already spent decades rendering Greek philosophy into Arabic (Successcribe, 2025). This essay's companion piece, "Why University, Of All Things," traces that distinction — independent invention against inherited idea — in full.

What these very different traditions shared was a starting orientation toward thought for its own sake — philosophy, theology, law, and the natural world, studied because understanding them was considered worthwhile in itself. The shift from school of thought to school of profession came later and more slowly, and it is from here that this essay narrows its focus to the Western university tradition specifically, since it is that tradition's particular professional degree structure — and its particular assessment methods — that generative AI is now disturbing most directly. Oxford's earliest teaching dates to around 1096, gaining real momentum once Henry II barred English students from Paris in 1167; Cambridge followed in 1209. From early on, both were structured around a foundational arts curriculum feeding into three parallel higher faculties — theology, law, and medicine — theology carrying the most prestige, but law and medicine present from the medieval structure's beginning rather than added later. Medicine, law, and eventually engineering, nursing, and social work were absorbed into the university's remit only as those professions themselves formalised, codified their own bodies of knowledge, and required a credentialing route into practice. It is this second, professional inheritance — not the ancient philosophical one — that the modern undergraduate degree in fields such as law, medicine, nursing, or social work has largely become. And it is this professional inheritance, specifically, that generative AI now unsettles.

Since professional degrees took their current form, the underpinning knowledge of a subject has been the thing the degree tested. A law student learned statute and precedent; a nursing student learned physiology and pharmacology; a social work student learned attachment theory and safeguarding legislation. The exam, the essay, the viva were, at root, a way of checking that this underpinning knowledge had taken root in the individual sitting the test.

Generative AI now performs the retrieval half of that task in seconds. Ask it for the relevant statute, the mechanism of action, the theoretical framework, and it arrives — adapted to the question, restructured to the context, ready to use. What has changed is not that the knowledge exists somewhere; it always did, in libraries and journals and the minds of lecturers. What has changed is the cost of reaching it, which has fallen to near zero. This is not a marginal concern or a plagiarism problem to be policed away: recent survey data shows the proportion of students directly including AI-generated text in assessed work rising sharply year on year, from 3% in 2024 to 12% in 2026, with generative AI use among undergraduates now close to universal rather than confined to a minority (HEPI and Kortext, 2026).

The deeper issue is definitional rather than statistical. Universities have long held a dual function: as places where learning happens, and as institutions that certify — through assessment — that a given individual has reached a given standard of competence. The rapid spread of generative AI unsettles that certification function specifically, by changing the conditions under which assessment evidence gets produced (MDPI, 2026). Where powerful AI tools sit a click away, a grade may increasingly reflect some blend of individual understanding and external cognitive support, rather than independent competence alone (MDPI, 2026). Whatever the essay used to measure, it can no longer straightforwardly be trusted to measure that same thing now.

None of this means the underpinning knowledge was pointless to teach, or that the crisis is really about AI at all. It means the test built to check for it has quietly stopped working — and everything that follows in this essay is an attempt to work out what ought to replace it, both inside the university and beyond its gates. This is not the first time the test has stopped working. It has happened before, more than once, and history is worth keeping close at hand as this essay proceeds — not as decoration, but as precedent for how an institution survives losing the thing it thought made it indispensable.

Section 2 — What Replaces Recall as the Thing Worth Testing

The university has faced this particular shape of crisis before. When Henry VIII broke from Rome and Thomas Cromwell's dissolution of the monasteries removed the religious houses embedded within Oxford and Cambridge in the late 1530s, the universities' near-exclusive claim to theological authority — the thing that had justified their existence for four centuries — could no longer be taken for granted. Cromwell's own injunctions of the period pushed the universities toward a broader humanist curriculum: readers in Greek and Hebrew, support for scholars beyond canon law, a widening of what "worth knowing" meant. The historian Diarmaid MacCulloch's account of this period makes clear it was neither smooth nor guaranteed — applications to Cambridge fell by roughly half in the years immediately following the dissolution, and a 1545 Act of Parliament came within a hair of liquidating the universities' own wealth for the Crown's benefit (MacCulloch, 2018). The institution survived not by defending its old monopoly but by finding something else worth being trusted to teach. That is the precedent worth holding onto here.

Section 1 left the underpinning-knowledge test broken without yet saying what should replace it. That question of what replaces student testing and assessment deserves a distinction drawn carefully, because the temptation is to collapse it. Underpinning knowledge was never only about recall. Learning a body of statute, or a set of physiological mechanisms, or a theoretical framework, did two things at once: it gave the student the specific facts, and it built — as something of a by-product — the mental scaffolding required to organise, question, and apply new information within that domain. Generative AI dismantles the first function almost completely. It does considerably less to the second, because scaffolding is not something that can be looked up; it has to be built, slowly, through the effort of holding ideas in relation to one another until that relation becomes second nature. A recent curriculum-reform paper puts the risk precisely here: what AI threatens is not routines but the slow work of reasoning, critique, and apprenticeship into disciplinary traditions, and automating foundational tasks without judgement risks the loss of both the practice and the underlying understanding of how knowledge is made (arXiv, 2025).

The problem is that most undergraduate assessment was never designed to separate these two things. An essay that demonstrates fluent command of a statute could, historically, only have been written by someone who had done the work of understanding it. That inference no longer holds, because the fluency can now be manufactured without the understanding ever having taken place. The assessment was a reasonable proxy for scaffolding for as long as recall and scaffolding were bundled together inside the same difficult task. Once recall becomes trivially cheap, the bundle comes apart, and the essay stops being evidence of anything beyond the ability to prompt a tool competently.

This reframes the question usefully. It is not "how do we stop students using AI," which is a policing problem with no durable technical solution. It is "how do we assess for scaffolding directly, now that we can no longer assess for it indirectly by testing recall." The more considered responses in this space have already begun moving in that direction — away from judging the polish of a finished written product, and toward judging the visible trail of reasoning, decision-making, and ownership that produced it: the working, not just the answer (Frontiers in AI, 2026): oral defence of written work, staged submissions that show a piece of thinking developing over time, and supervised tasks that put a student's judgement under direct observation rather than inferring it from a document that could have been produced by anyone, or anything.

None of this is costless. Assessing process rather than product is more expensive to run, harder to standardise across a cohort of hundreds, and less amenable to the anonymous, arm's-length marking that has underpinned undergraduate assessment at scale for decades. There is also a weight to it that recall-based testing never carried: judgement, unlike a fact, has an ethical dimension built in — assessing whether a student can hold that weight responsibly is a different and harder task than assessing whether they remember the statute, and it is a task this essay returns to at its close. But the alternative — continuing to mark essays as though they still reliably indicate what they used to indicate — is not really an alternative. It is a slow accumulation of degrees whose grades mean less than they did, awarded by institutions that have not yet admitted this to themselves, in a milder echo of the same institutional reluctance that delayed the Tudor universities' own reckoning with what they had actually become.

Section 3 — Degree Apprenticeships as a Structural Answer, With Limits

If assessment needs to move toward observing judgement directly, one obvious place to do that is where judgement is already being exercised for real: in the workplace, on live cases, with real consequences attached to getting it wrong. This is the strongest argument for degree apprenticeships, and it is a structural one rather than a matter of curriculum tweaking. The claim here is not that apprenticeships add useful practical experience to a degree that otherwise stays the same. It is that the informal, unwritten system that used to complete a graduate's formation — the mentorship, correction, and slow acquisition of professional judgement that happened after graduation, on the job, under supervision — has quietly stopped functioning reliably, for reasons largely unconnected to AI, and that embedding this formation inside the degree itself is a response to that collapse rather than an enhancement bolted onto an otherwise-unchanged curriculum.

This is not a new idea competing with an old, settled tradition; it is closer to a return. The redbrick civic universities — Manchester, Birmingham, Liverpool, Leeds, Sheffield, Bristol, and others — were founded in the nineteenth century explicitly to serve the industrial cities around them, embedding practical and professional relevance into higher education from their outset rather than treating it as an afterthought to a classical curriculum. The historian William Whyte's account of these institutions argues they were, in their own moment, the biggest influence on the shape of new university foundations that followed (Whyte, 2015). Degree apprenticeships are not a break from university tradition so much as a rediscovery of the civic model's founding logic, dressed in contemporary form.

Degree apprenticeships are not a hypothetical fix in the present either. They are already a substantial and growing part of the UK postsecondary landscape, accounting for more than 6% of new postsecondary enrolments, and international comparisons — Germany's dual-training system, at around 5% of higher education enrolment — show a version of this integration operating at national scale rather than as an experimental pilot (Work Shift, 2025). Where they work well, they work because the profession itself is codifiable: nursing, engineering, construction, and much of social work practice have a defined, observable skillset that can be assessed in situ, under supervision, against a recognisable standard. A workplace assessor can watch a trainee nurse manage a medication round or a trainee social worker conduct a home visit and form a judgement about competence that no essay could reliably replicate. A 2026 paper on post-AI apprenticeship goes further, arguing the problem has been misdiagnosed as a curriculum-content issue when the real cause is structural: the informal, post-degree apprenticeship system that historically completed graduate formation no longer reliably exists, and proposing embedded, structural redesign rather than syllabus revision as the proper response (Journal of Higher Education Theory and Practice, 2026).

Credit is due here to a pattern already visible in practice: many university departments — social work courses among them — deliberately staff their teaching with former practitioners, lecturers who have themselves done the safeguarding visit, sat in the multi-agency meeting, carried the caseload. This solves a real problem: an assessor without that grounding cannot reliably judge professional judgement in someone else, however well they understand the theory behind it. Departments that have built practitioner expertise into their teaching staff deserve to be credited for taking that seriously rather than assuming subject knowledge alone qualifies someone to assess it.

But expertise and impartiality are separate problems, and solving the first does not touch the second. Moving assessment toward oral defence, staged discussion, and supervised observation — as Section 2 argues it should — shifts the locus of judgement away from an anonymous marking scheme and onto an individual assessor sitting in a room with a specific student they may have taught for three years and likely have a personal relationship with. A blind-marked script is at least partially insulated from that relationship. A viva is not. This essay's companion piece, "Academia and the Marketplace," has already named the mechanism in a different context — the temptation to grade generously when a programme's continuation depends on student satisfaction and retention — and the same pressure applies directly here: the more an assessment depends on an assessor's discretion, the more exposed it becomes to sympathy, fatigue, personal rapport, or simple reluctance to be the reason a struggling student fails. Practitioner grounding makes an assessor more qualified to judge well. It does nothing to guarantee they will judge without favour. This is not an unaddressed gap in principle: UK universities already run exactly this kind of safeguard for traditional written assessment, through external examiners who dip-sample internally marked work each year under UK Quality Code moderation requirements, with social work qualifying programmes carrying an additional layer of oversight from the profession's own regulators. The open question is not whether such safeguards exist, but whether they scale to the kind of assessment Section 2 is proposing — a viva, a staged discussion, a supervised observation — in the way they scale to a marked script. Sampling a batch of essays after the fact is a fundamentally easier task than sampling a live conversation for the presence or absence of sympathy, and a serious version of process-based assessment needs to extend existing moderation practice into that harder territory specifically, rather than assuming the safeguards built for one form of assessment transfer cleanly to another.

This is also precisely where the model's limits show. Not every discipline has a single, fixed workplace destination against which judgement can be observed and assessed. Pure sciences, much of the humanities, and the more theoretical ends of social sciences do not have an equivalent of "the ward" or "the home visit" — a bounded, repeatable, observable setting in which competence declares itself. For these disciplines, degree apprenticeships are not simply harder to design; they may be the wrong tool altogether, because the thing being taught was never primarily a workplace skill in the first place. Institutional response has, so far, lagged behind this diagnosis considerably: most institutions have framed the AI challenge in terms of plagiarism and academic integrity policy rather than assessment redesign, despite general recognition that teaching and assessment ought to reflect the skills an AI-influenced landscape actually requires (Frontiers in Education, 2026). Whatever replaces recall-based assessment in those disciplines will need to be found elsewhere, and this essay does not claim to have that answer. What can be said with more confidence is that for the professional degrees this essay is concerned with — the direct descendants of medicine and law rather than of philosophy — structural integration with the workplace is not an optional enrichment. It is close to a necessity, provided the assessors doing the observing are held to the same rigour they are meant to be applying.

Section 4 — Evolution Reclaims Its Path

There is a way of reading the last twenty-five years of UK higher education that treats 1999 not as evolution but as intervention. Tony Blair's 50% participation target, and the fee-funded machinery built to sustain it, did not emerge from organic market pressure the way a genuine evolutionary change would. It was injected directly into the system by a single political act — closer to genetic engineering than to natural selection, redirecting several hundred years of a slower academic tradition toward a mass-credentialing model within a single generation. The historian Peter Mandler's account of Britain's postwar transition to mass education places 1999 as a culmination rather than a bolt from nowhere — the pressure toward expansion had been building steadily across the second half of the twentieth century, through the Robbins Report of 1963 and beyond, before finally reaching the scale of a formal 50% target (Mandler, 2020). Whatever its stated purpose in widening opportunity, the structural effect was to convert large parts of the university sector into something closer to a commercial enterprise: departments justified by recruitment numbers, courses judged by employability data, research prioritised by fundability rather than by the discipline's own internal logic. Sheila Slaughter and Gary Rhoades gave this pattern a name in 2004 — academic capitalism — describing the systematic refashioning of knowledge as commodity and students as consumers across the sector (Slaughter and Rhoades, 2004). Christopher Newfield's later account of the same drift argued that privatisation of this kind actually impoverishes the institutions it claims to strengthen, leaving them as debt-financed shells of their former purpose (Newfield, 2008). The mechanism traced here is the same one this essay's companion piece, "Academia and the Marketplace," follows in detail through its 2025–26 financial consequences.

Genetically engineered systems carry a particular risk that naturally evolved ones do not: the engineered trait can be reclaimed, unwound, or exposed as unstable the moment conditions change, precisely because it was never selected for by the pressures actually operating on the system. Generative AI may be exactly this kind of pressure. It has no capacity for philosophical contemplation, no stake in ambiguity, no ability to sit with a question that has no resolution — the qualities the ancient schools of thought in Section 1 were built around, before profession and credential were ever part of the picture. What it can do, and do very well, is exactly what the mass-credentialing model asks of it: transmit codified information efficiently, on demand, at near-zero cost. Philosopher Lily Abadal makes this case directly: the sort of knowledge that is embodied, and requires the mastery of tools or the correction of a supervising hand, is far less vulnerable to AI substitution than the sort of knowledge that consists of efficiently transmitted information, however prestigious the institution transmitting it happens to be — and, crucially, this holds regardless of whether the discipline in question is vocational or academic, since any programme that has reduced itself to information transfer and credentialing is equally exposed (Abadal, 2025). On this reading, the departments and disciplines most exposed to AI disruption are not defined by subject matter — arts against sciences, humanities against vocational — but by function: whichever parts of the university most completely converted themselves into information-transmission and credentialing operations over the last quarter-century are the parts most directly exposed now, regardless of what subject sits on the nameplate. This is evolution reclaiming a path that was diverted rather than chosen — the genetically engineered trait proving unstable precisely where the underlying pressure it was built against reasserts itself.

This is not the only reading in circulation, and the loudest counter-argument belongs alongside it rather than past it. Palantir's Alex Karp has argued the opposite conclusion at some volume: that AI will "destroy humanities jobs" specifically, while vocational and technical training survive and even flourish, framing the divide in explicitly subject-matter terms — philosophy graduates exposed, welders and electricians safe (Karp, Davos remarks, January 2026; TBPN podcast, March 2026). It is a reasonable prediction on its own terms, resting on real observed trends in entry-level hiring. But it rests on two assumptions that deserve naming rather than granting quietly. First, it assumes AI's current trajectory continues largely unchecked into capabilities it does not yet reliably possess — a forecast dressed as an observation. Second, it assumes subject-matter is the correct axis of vulnerability at all, when the function-based reading above suggests otherwise: a humanities course that still teaches genuine, uncertain, effortful inquiry may be considerably safer than a vocational course that has itself been reduced to credentialed information transfer. Karp's own commercial position matters here too — Palantir has publicly criticised universities for "indoctrinating" students and operates its own vocational-style alternative credentialing route (its Meritocracy Fellowship), which does not make the prediction wrong, but does mean it is not disinterested. A prediction delivered with this much rhetorical force, tied explicitly to which political tribe gains or loses economic power as a result, has drifted some distance from economic forecasting into something closer to prophecy — compelling to listen to, but properly treated as argument rather than as settled fact.

Section 5 — The Labour Market Context

None of this happens in a vacuum. Evidence points to an environment that complicates the apprenticeship argument rather than simply supporting it. Entry-level hiring across several sectors — most visibly technology, but not confined to it — has contracted sharply over the past two to three years: UK tech graduate roles fell 46% in 2024, with a further 53% drop projected by 2026, and US entry-level tech postings are down roughly 30–35% since early 2023 (FastApply, 2026; Rezi.ai, 2026). Multiple independent data sources, built on different methodologies and produced by organisations with no obvious shared interest in the finding, converge on the same underlying pattern: the contraction is concentrated specifically at the entry point, among the youngest cohort of workers, rather than spread evenly across the workforce — software developers aged 22–25 are down nearly 20% from 2024 peaks (Digital Applied, 2026, citing Stanford AI Index 2026 and an Anthropic economic study). Graduate unemployment has, on some measures, overtaken general unemployment for the first time since records began — a genuine reversal of the historic graduate wage and employment premium (Metaintro, 2026). This is not a distant academic debate: the literary historian Stefan Collini's own commentary on the sector, published this June, gives a clear-eyed diagnosis of universities' deteriorating financial position, arguing the fee-dependent model built from 1999 onward is being eroded from multiple directions simultaneously (Collini, 2026) — scholarship and lived crisis converging in real time rather than one following the other at a comfortable historical distance.

Consequently, a caveat belongs here rather than being smoothed over. At least one credible study finds that the shift to remote working predicts this decline in entry-level hiring at least as well as AI adoption does, on the reasoning that remote arrangements raise the cost of the informal supervision and correction that used to happen simply by junior staff sitting near senior staff (Lambert and Schindler, 2026). This does not overturn the AI-driven explanation, but it does mean the causal picture is genuinely contested rather than settled, and an essay that treats "AI did this" as a closed question is overstating what the evidence currently supports. What both explanations share, notably, is a common mechanism: something has made the informal, low-cost supervision of junior staff more expensive or less available, whether that something is a screen full of AI tools or an empty office.

Whichever cause dominates, the effect for this essay's argument is the same. If the traditional route into a profession — get a degree, then learn the rest on the job — is narrowing regardless of the exact mechanism, then a structural response that brings supervised, judgement-based learning forward into the degree itself is not simply a good idea in principle. It is closer to a necessary adaptation to conditions that already exist and show no sign of reversing. What employers report wanting from the graduates who do get hired reinforces this: less emphasis on theoretical knowledge that can now be generated on demand, and more emphasis on demonstrated experience, applied judgement, and the kind of critical thinking that sits underneath any particular tool or technique — the World Economic Forum's Future of Jobs Report 2025 places analytical thinking at the very top of globally demanded skills, framed explicitly as AI-resistant (World Economic Forum, 2025) — precisely the capabilities that Section 2's process-based assessment and Section 3's workplace-embedded learning are designed to build and evidence.

Section 6 — The Reverse Obligation

There is a temptation, at this point in the argument, to conclude that the burden of adaptation falls entirely on universities — that it is for higher education to redesign its assessment, restructure its degrees, and integrate more closely with industry, while employers simply wait to receive better-prepared graduates. This would be a mistake, and not merely for reasons of fairness. It would also fail to solve the actual problem, because much of what is currently breaking down is not a university failure at all. It is the erosion of a shared arrangement in which employers held up their side.

For decades, the formation of a competent professional was genuinely a joint undertaking. Higher education supplied foundational knowledge and analytical grounding; employers supplied context, correction, and the accumulated tacit knowledge that only comes from doing the job under supervision, and entry-level roles were the bridge connecting the two (NACE, 2026). That bridge depended on employers being willing to absorb the cost of training people who were not yet fully productive — a cost that has become harder to justify as technical skills date faster, margins tighten, and AI tools raise the bar for what "productive from day one" is assumed to mean. The result is a coordination failure rather than a one-sided one: if every employer simultaneously wants graduates who arrive fully formed while declining to invest in forming them, the collective system that used to produce fully formed professionals stops working for everyone, including the employers who are, individually, behaving rationally (NACE, 2026).

There are already voices from inside industry making this argument rather than only from within higher education, which matters, because it means the case for reciprocity is not simply universities protecting their own position: one UK employer-side voice argues plainly that businesses have a responsibility to invest in entry-level training instead of expecting graduates to arrive fully formed, since technical skills age quickly anyway and what actually lasts is curiosity, judgement, and adaptability (TechRound, 2026, quoting Victoria Knight, Node4). The scale of the strain is visible in the numbers: 81% of employers now expect to prioritise demonstrated work experience over completed credentials when assessing candidates, even as shrinking entry-level hiring makes that experience structurally harder to obtain (World Economic Forum, 2026), and a parallel UK survey finds 57% of employers reporting skills shortages while 43% simultaneously hired fewer staff over the past year (People Management, 2026). The proposed fixes in this space point in a consistent direction: co-op models, paid structured work placements embedded in degree progressions, and formal industry partnerships that treat the supervision of junior staff as a shared cost rather than one party's problem to solve alone (Strada Education Foundation, 2025). These are, in effect, the employer-side mirror of the degree apprenticeship structures discussed in Section 3 — the same integration, approached from the other direction, and closer in spirit to the redbrick civic bargain than to anything 1999 produced. Neither half works well without the other. A university that redesigns its degrees around workplace-embedded judgement, but finds no employers willing to provide the workplace, has simply moved the bottleneck rather than removed it. An employer sector that demands demonstrated experience while continuing to shrink the entry-level roles that used to provide it is asking for an output it has stopped funding the input for.

Close — What's Left for the University

Picture a defendant standing in the dock in 2032. Opposing counsel, the judge, and the press in the gallery all have a tablet open in front of them, each running an AI agent that can, in the seconds after the defendant finishes speaking, generate a sharper counter-argument than any human in the room could produce unaided — cross-referencing every precedent, every inconsistency, every weakness in what has just been said, faster than the next sentence can be formed. Whatever underpinning legal knowledge that defendant, or their barrister, might have prepared for this moment does not compete with that. The retrieval war is lost before it begins, and it is worth pausing on what that actually means for the person standing there: not an abstraction, but someone whose liberty depends on an argument that machinery in the room can already out-analyse.

This is not really a scene about courtrooms, and it is not far off. It is what Section 1's argument looks like once it stops being abstract. If a written submission can no longer be trusted as evidence of what a person knows, then a spoken one, tested in real time against an AI opponent, is even less of a contest — knowledge alone simply cannot hold the room. Section 4 argued that embodied, workplace-based judgement resists AI substitution better than recall-based knowledge does; this scene is the reminder that the protection is not absolute, because AI does not have to replace the judgement itself to change the contest — it only has to sit in the hands of whoever is arguing against that judgement. And yet something in that courtroom still has to decide the case. Something still has to weigh the argument the AI produced, the argument given in return, the history of the person standing in the dock, and the question of what a just outcome actually looks like for everyone affected by it — not merely which side assembled the more comprehensive brief. Whether that outcome bends toward retribution or toward something closer to restorative justice is not a question any AI agent in that gallery has a stake in answering. It is a human question, asked by a human judge, about what should happen to another human being. The university did not start as a school of profession; it became one, slowly, as medicine, law, and eventually the newer professions absorbed into its remit and required it to certify competence rather than only cultivate thought. That professional inheritance is the part generative AI is now disturbing, by making the traditional evidence of underpinning knowledge — the fluent essay, the well-structured exam answer, the fluent argument in a courtroom — no longer reliable proof of anything in particular.

None of this is a single story, and it is worth naming the whole shape of it once, here, at the end. Ancient schools of thought became medieval schools of profession. The Tudor dissolution stripped away the universities' theological monopoly and forced a genuine widening of what counted as worth teaching. The Victorian civic universities rebuilt higher education around industrial relevance from the ground up. The postwar decades built the pressure that 1999 finally converted into mass, fee-funded expansion — injected rather than evolved, Section 4 argued, which is precisely why it has proved unstable under new pressure. And now generative AI is doing to recall-based assessment what each of those earlier crises did to whatever came before it: reclaiming a path that a single political act diverted, rather than imposing something genuinely new. The university now observed is not the product of one founding moment; it is the accumulated residue of every crisis it has been forced to survive, each one leaving the institution different from what it was before, and each one, on the evidence of the last, eventually finding something worth being trusted to do next.

What is left for the university, in this professional register at least, is not less than it had before, but different. That distinction cuts against the subject-matter framing offered earlier by voices such as Karp's: the vulnerable programmes are not defined by which faculty they sit in, but by whether they still ask a student to do the harder, slower work at all — a vocational course reduced to credentialed information transfer is no safer than a humanities course that has done the same, and a humanities course that has kept effortful, uncertain inquiry alive is no more exposed than the trades Karp names as secure. It is no longer primarily the custodian of retrievable knowledge; that function has been substantially, and probably permanently, ceded to tools that perform it faster and more cheaply than any lecture, or any barrister's brief, ever could. What remains properly the university's to do is the harder, slower work: building the scaffolding that lets a graduate use retrieved knowledge with judgement rather than simply access it, and doing so in settings — increasingly workplace-embedded ones — where that judgement can actually be observed rather than merely inferred from a polished document. For the professions this essay has been concerned with, that work now has an explicit ethical dimension attached to it that recall-based training never had to carry on its own: if AI can win the argument, diagnose the mechanism, or draft the care plan, the university's remaining task is to form the person who decides, wisely and humanely, what should be done with that answer once it has been produced — whoever that person turns out to be, judge, doctor, nurse, or social worker, wherever a judgement, and not just an answer, is required.

None of this can be accomplished by higher education acting alone, and the essay's final claim is the least comfortable one for either side to hear in isolation. Universities cannot assess judgement they have no access to observing, and industry cannot receive judgement it has stopped investing in developing. The old school of thought asked, for its own sake, what was worth knowing. The school of profession that grew out of it now has to ask a harder question together with the industries it serves: not what is worth knowing, but what is worth learning to do, and who is prepared to stay in the room long enough to teach it.

References

Abadal, L. M. (2025). Only the Humanities Can Save the University from AI. Public Discourse, 21 July 2025.
Ancient Universities of India. Wikipedia (2025).
Chirikov, I. et al. (2026). Generative AI Use and Misuse Call for Assessment Reform in Higher Education. Science.
Collini, S. (2026). Squadrons of Pigs: Bonfire of the Universities. London Review of Books, Vol. 48, No. 11, 25 June 2026.
Designing AI-Resilient Assessment in Higher Education: A Four-Pillar Conceptual Framework. Frontiers in AI, Vol. 9 (2026).
Digital Applied (2026). AI and Jobs in 2026: What the Labor Data Really Shows, citing Stanford AI Index 2026 and an Anthropic economic study.
Dukes Plus (2025). What Is the Oldest University in the World?
Musacchio, F. (2025). Hellenistic Encounters with India: Historical and Cultural Background. Independent historical overview, on Indo-Greek and Greco-Bactrian cultural fusion.
FastApply (2026). Entry-Level Jobs Are Disappearing: What New Grads Need to Do Differently in 2026, citing World Economic Forum Future of Jobs Report 2025.
A Framework for AI and Vocational Assessment in Higher Education. Frontiers in Education (2026).
HEPI and Kortext (2026). Student Generative AI Survey 2026. HEPI Report 199.
Journal of Higher Education Theory and Practice (2026). Apprenticeship after AI: Bridging Gaps in Early-Career Knowledge-Work Roles.
Karp, A. (2026). Remarks in conversation with Larry Fink, World Economic Forum Annual Meeting, Davos, January 2026.
Karp, A. (2026). Remarks on TBPN podcast, March 2026.
Lambert, P. J. and Schindler, Y. (2026). The Broken Ladder: AI, Remote Work, and Early-Career Hiring. London School of Economics working paper, with Warwick and Oxford Ellison Institute co-authors.
List of Oldest Higher-Learning Institutions. Wikipedia (2025).
MacCulloch, D. (2018). Thomas Cromwell: A Life. Allen Lane.
Mandler, P. (2020). The Crisis of the Meritocracy: Britain's Transition to Mass Education since the Second World War. Oxford University Press.
MDPI (2026). Assessment Validity in the Age of Generative AI: A Natural Experiment. Special Issue: Generative AI in Higher Education.
Metaintro (2026). 5.7% and Climbing — What the 2026 Job Market Means for New Graduates.
NACE (2026). Op-ed: If Employers No Longer Train Entry-Level Hires, What Is Higher Education's Role?
Newfield, C. (2008). Unmaking the Public University: The Forty-Year Assault on the Middle Class. Harvard University Press.
People Management (2026). New Report Warns More Support Is Needed for Young People Entering Work. Business Barometer 2026.
Pumbedita Academy. Wikipedia (2025).
Rezi.ai (2026). The Crisis of Entry-Level Labor in the Age of AI (2024–2026), citing NACE Class of 2026 data.
Slaughter, S. and Rhoades, G. (2004). Academic Capitalism and the New Economy: Markets, State, and Higher Education. Johns Hopkins University Press.
Strada Education Foundation (2025). Employer and Community College Partnerships: Models and Practices.
Successcribe (2025). Top 20 Oldest Universities in the World.
Sura Academy. Wikipedia (2025).
TechRound (2026). World Youth Skills Day 2026: Are Universities Teaching Students Skills That Employers Actually Need?, quoting Victoria Knight, Node4.
The Chopras (2023). Top 10 Oldest Universities in the World You Should Know About.
UK Standing Committee for Quality Assurance (UKSCQA) (2018). UK Quality Code for Higher Education: External Expertise.
Whyte, W. (2015). Redbrick: A Social and Architectural History of Britain's Civic Universities. Oxford University Press.
Work Shift (2025). Reporting on German degree apprenticeships and the Reach University model.
World Economic Forum (2025). Future of Jobs Report 2025.
World Economic Forum (2026). Entry-Level Hiring Is Tougher Than Ever. What's Helping Grads Now.
arXiv (2025). AI Education in Higher Education: A Taxonomy for Curriculum Reform and the Mission of Knowledge. Preprint 2509.23363.

See also: Academia and the Marketplace (YoungFamilyLife, 2026) for the UK-specific financial and policy detail underlying the 1999 expansion and its consequences, referenced throughout Sections 3, 4, and 6.

Topics: #HigherEducation #AIAndEducation #UniversityFunding #GraduateEmployment #DegreeApprenticeships #Marketisation #IndependentEnquiry