For at least a century, formal education has bundled together several functions that do not naturally belong together. It teaches, sorts, certifies, socialises and, in some cases, shelters young adults from the labour market for a few years longer. Generative AI is not simply improving one part of that arrangement. It is exposing the weakest compact inside it: the assumption that a written artefact, produced largely out of sight, is a reliable proxy for understanding.
The immediate panic about cheating missed the deeper institutional point. Once machine systems can draft essays, summarise literature, generate code, solve routine problem sets and mimic professional tone, the cost of producing plausible evidence of competence falls sharply. That does not mean expertise disappears. It means the evidence traditionally used to infer expertise becomes less trustworthy. Education, especially higher education, is therefore being drawn into a new business whether it likes the term or not: auditing judgment.
The end of scarcity in plausible output
For much of modern schooling, institutions treated polished output as scarce. A coherent essay, a correct derivation, a competently structured report or a passable software script suggested not only knowledge but effort, organisation and some degree of independent thought. That logic depended on production being difficult. AI has altered the economics. Plausible output is now abundant.
Large language models do not possess understanding in the human sense, but they are effective engines for producing forms that look like understanding. As research and policy reviews from UNESCO, the OECD and national education authorities have noted, this changes incentives across the system. Students can offload drafting. Teachers can automate feedback. Employers can no longer assume that a portfolio of elegant prose or neat code was created in conditions that reveal an individual’s own capability.
The scarce asset is no longer access to information, but credible proof of how a person uses it. That is a more demanding problem than plagiarism detection. It requires educational institutions to observe performance in ways that are richer, more contextual and harder to counterfeit.
Why the diploma is not disappearing
A common prediction is that AI will dissolve the degree altogether, replacing it with fluid skills markets and fine-grained micro-credentials. That is too simple. Diplomas persist not because they perfectly measure competence, but because they compress uncertainty. Employers use them as shorthand for persistence, selection, social fit and baseline literacy, even when they know the signal is noisy.
What is changing is the meaning of the diploma. The degree is moving from a signal of completed study to one layer in a thicker record of observed performance. In an AI-saturated environment, institutions that can supplement broad certification with trustworthy evidence of judgment will hold their value best. Those that rely mainly on unsupervised written coursework may find that their credentials still circulate, but with a growing discount applied by employers and professional gatekeepers.
This does not imply a return to old-fashioned closed-book examination as the universal answer. In many occupations, the use of AI tools will be normal, just as calculators, search engines and software environments are normal. The question is not whether assistance is allowed. It is whether assessment can distinguish between tool use that extends judgment and tool use that replaces it.
From testing memory to testing decisions
The curriculum debate is often framed around what facts students should still memorise when AI can retrieve and rephrase so much information. Useful as that question is, it understates the shift under way. Assessment is becoming less like inspection of stored knowledge and more like an audit of decision-making.
The scarce asset is no longer access to information, but credible proof of how a person uses it.
Consider what matters in professional practice. A clinician must decide whether a generated summary omits a contraindication. A civil servant must notice when a briefing is fluent but based on weak evidence. An engineer must recognise when generated code is operational but insecure. A historian must detect when a persuasive narrative has compressed ambiguity into false certainty. In each case, the task is not mere recall. It is situated judgment under conditions of cognitive outsourcing.
That suggests a different design for education. Institutions will need more assessments in which students must justify choices, document how they used tools, respond to adversarial prompts, revise work after challenge and explain failures. The object being tested is not whether a student can produce a pristine artefact at first pass, but whether they can reason their way through uncertainty, check machine output and remain accountable for the result.
Assessment is becoming less like inspection of stored knowledge and more like an audit of decision-making.
The rebirth of oral, practical and live assessment
One consequence is the quiet return of older forms. Oral examination, studio critique, laboratory demonstration, supervised simulation and viva-style defence all look newly attractive, not because they are nostalgic, but because they reveal thinking in motion. They make it easier to probe whether a student can explain assumptions, defend trade-offs and recover when challenged.
These methods are expensive. They require trained staff, time and clear rubrics. They also create risks of inconsistency and bias if poorly designed. Yet they answer a problem mass written assessment no longer solves well. A live encounter can establish whether competence is robust or merely staged.
Professional schools have long understood this. Medicine uses clinical placements and observed encounters. Architecture uses crits. Doctoral education uses the viva. The novelty is that similar logics are likely to move outward into domains that once depended more heavily on take-home writing and standardised responses. In the AI age, institutions may have to spend more to assess less volume but better evidence.
The politics of friction
There is, however, a political difficulty. Students and policymakers have been promised for years that digital education would become more scalable, more personalised and more efficient. Judgment-rich assessment moves in the opposite direction. It introduces friction. It is slower, costlier and less easily automated.
That friction may nevertheless be the price of legitimacy. When the surrounding information environment is flooded with synthetic fluency, institutions gain authority not by accelerating throughput but by making evaluation more discriminating. The value of the institution lies partly in its willingness to impose procedures that are inconvenient precisely because they produce more credible signals.
This will be uncomfortable for systems already under budget pressure. Yet the alternative is worse: a gradual hollowing-out in which grades remain plentiful, transcripts remain tidy, but confidence in what either signifies steadily declines.
Curriculum after the age of solitary authorship
The degree is moving from a signal of completed study to one layer in a thicker record of observed performance.
If assessment changes, curriculum cannot remain untouched. The old model assumed a substantial zone of solitary authorship in which students converted reading into original written performance. That model will survive in some settings, but it is no longer the default environment in which knowledge work occurs. Most graduates will operate with persistent machine assistance.
Curricula should therefore teach three layers at once. The first is domain knowledge: concepts, methods and standards without which no critical use of tools is possible. The second is procedural fluency with AI systems: prompting, verification, decomposition of tasks and awareness of limitations. The third, and most neglected, is epistemic governance: knowing when evidence is weak, when outputs are unverifiable, when a task requires human consultation and where responsibility lies if decisions go wrong.
This third layer is what many institutions still treat as tacit. Yet it is increasingly the core of professional competence. A graduate who can produce polished output with machine assistance but cannot judge evidential quality is not future-ready. They are merely well equipped to make high-speed mistakes.
The unequal burden on knowledge institutions
Not all institutions will adapt equally. Elite universities, selective professional schools and well-funded technical institutes can afford smaller-group assessment, richer supervision and stronger integrity processes. Mass systems with limited staffing face a harsher dilemma. They are more exposed to low-trust assessment, yet have fewer resources to redesign it.
This could intensify educational stratification. The privileged may receive the kind of education that still offers close observation of reasoning, while everyone else is processed through cheaper modes that are easier to game. If so, the real divide in the AI age will not simply be access to tools. It will be access to institutions capable of witnessing and certifying genuine judgment.
That danger should concern public systems in particular. The legitimacy of mass higher education rests on the claim that it can confer not only opportunity but trustworthy standards. If AI causes standards to become more opaque, the social contract around public credentials weakens.
The degree is moving from a signal of completed study to one layer in a thicker record of observed performance.
Employers will not solve this alone
Some argue that employers will simply bypass educational credentials, using work samples, probation periods and task-based hiring instead. Elements of that shift are real, especially in digital occupations. But firms face the same authenticity problem as universities. A take-home exercise can be heavily AI-assisted. A polished portfolio can be machine-shaped. Even interviews are vulnerable to coaching and generated preparation.
Employers can evaluate performance in context, but they cannot cheaply replicate the full developmental and auditing role of education. Nor should they wish to. If every firm must build its own high-resolution assessment apparatus, labour markets become less portable and more exclusionary. Shared institutions remain necessary to create common, trusted signals.
The likely outcome is not disintermediation but renegotiation. Employers will ask more of educational credentials, professional bodies will refine licensure and universities will be pressed to show not merely that students completed work, but under what conditions and with what demonstrated level of independent judgment.
Assessment is becoming less like inspection of stored knowledge and more like an audit of decision-making.
The rise of process evidence
One underappreciated consequence of AI is that process may matter more than product. A final answer reveals less than a record of intermediate decisions: what sources were consulted, which outputs were rejected, what assumptions changed and how feedback altered the work. In fields from programming to design to policy analysis, these traces can show whether a learner understood the terrain or simply accepted the first plausible result.
Educational institutions will therefore be tempted to collect more process data. That has promise, but also limits. Surveillance-heavy learning environments may improve verification at the cost of trust and autonomy. NIST’s work on AI risk and UNESCO’s ethical guidance both point to the need for proportionality, transparency and governance. Not every educational problem should be solved by more monitoring.
The better route is selective evidencing: requiring students to expose reasoning where it matters, rather than treating every keystroke as an object of inspection. The aim is not total visibility. It is accountable demonstration.
What remains human, and why that is not enough
Educational rhetoric often replies to AI by celebrating uniquely human qualities: creativity, empathy, ethics, curiosity. The sentiment is understandable, but as institutional guidance it is weak. Many failures in AI-rich environments will not arise because people lack humanity. They will arise because they fail to exercise disciplined professional judgment in tandem with tools.
Creativity without verification can become stylish error. Empathy without evidential rigour can become misguided confidence. Curiosity without methodological restraint can produce noise. What institutions must certify is not a generic residue of humanness, but a capacity to act responsibly in systems where cognition is distributed across people, software and organisations.
That is a more exacting ideal than the romantic defence of the human. It asks education to produce adults who can remain answerable even when they did not generate every line, every paragraph or every analytical step themselves.
An old mission in a harsher form
In one sense, nothing here is new. Universities and professional schools have always claimed to cultivate judgment, not just transfer information. The difference is that AI is forcing them to prove it. When machine systems can imitate many surface features of expertise, institutions can no longer rely on artefacts alone to certify the person behind them.
The coming audit of expertise will not abolish the lecture, the textbook or the degree. It will, however, make visible which institutions are genuinely able to observe reasoning, test accountability and defend standards in public. Those that can do so will remain central to the knowledge order. Those that cannot may still issue credentials, but with diminishing authority.
Education’s prestige has long rested partly on scarcity: scarce places, scarce knowledge, scarce access to credentialed status. In the AI age, the more relevant scarcity is credible evaluation. The institutions that matter will be the ones able to say, with defensible confidence, not merely that a student completed a course, but that they demonstrated judgment when fluent machine help was available and the easy answer was close at hand.
