Verifiable judgment: what AI actually demands of universities
This blog was kindly authored by Mauricio G. Villena, Dean and Professor of Economics and Public Policy, Diego Portales University, Chile.
Generative AI is most often discussed in universities as a threat to academic integrity. That framing captures something real, but it misses the more important question. According to HEPI and Kortext’s 2026 Student Generative AI Survey, 94% of UK undergraduates now use AI for their assessed work – up from 88% the previous year. The issue is not primarily that students are using these tools. It is what that usage reveals about the assessments themselves. If so much assessed work can be produced by a machine, the assessment was probably not measuring what we thought.
This is a structural problem and it predates AI. Universities have long assessed students on outputs – essays, reports, examinations – that are proxies for underlying capabilities rather than demonstrations of them. AI has made the gap between proxy and capability visible and urgent. The question now is whether institutions will respond by redesigning their assessments or by investing in better detection tools. Detection is a losing strategy. The response must be redesign.
What should assessment measure?
The answer is what might be called verifiable judgment: the capacity to evaluate evidence critically, to reason under conditions of genuine uncertainty, to write with precision and accountability, and to apply knowledge in contexts that are not scripted in advance. A student who produces a competent essay with AI assistance has not demonstrated that capacity. A student who can identify where the AI’s reasoning fails, construct a better argument, and defend it under questioning has begun to.
This implies a significant shift in how learning is assessed and accredited: toward oral examinations, iterative project work, supervised problem-solving, and portfolios reviewed against explicit criteria. The Office for Students should make AI assessment policy an auditable condition of provider registration, not guidance that institutions can interpret at will. The Quality Assurance Agency should update its threshold standards to require a minimum proportion of assessment that is demonstrably non-delegable. These proposals are extensions of existing quality assurance logic to a changed environment.
The credential architecture requires parallel attention. The proliferation of micro-credentials, stackable qualifications, and competency-based certifications is welcome in principle: people should be able to build capabilities in stages, compatible with working lives. But a credential creates value only when it is anchored in well-defined learning outcomes and supported by quality assurance mechanisms robust enough to make it portable and legible. Without that infrastructure of trust, modularity produces a proliferation of badges that employers cannot interpret and that students cannot rely on. The government’s Lifelong Learning Entitlement must make transparent, verifiable learning outcomes a legal condition of public recognition.
There is a further dimension that tends to be lost in the urgency of the AI debate. Sustainability is not a single concept in this context. There is education for a sustainable world – equipping graduates to address climate change, inequality, and structural transformation. And there is education that sustains individuals over time, providing capabilities that allow people to navigate a world in which the skills demanded will continue to change across the length of a career. Universities that take both seriously will need to integrate these purposes into their curriculum design. The risk – and it is a real one – is that the emphasis on employability and lifelong learning becomes purely instrumental if disconnected from purpose and citizenship. Skills without judgment, and judgment without values, are insufficient responses to the challenges graduates will face.
In the long run, universities do not compete on who digitalises fastest or who launches the most programmes. They compete on credibility – the belief, held by students, employers, and governments alike, that a qualification means what it says it means. Rebuilding that credibility in the age of AI requires not better marketing or more sophisticated detection. It requires redesigning the conditions under which verifiable judgment is actually formed, certified, and recognised.
That is the agenda. It is more demanding than the policy debate so far has acknowledged.
Comments
Paul Vincent Smith says:
It’s interesting, in only the fifth year of widespread AI use, to see how quickly the memes and shibboleths surrounding the position of AI in assessment have become established. First, the positives: Prof. Villena successfully outlines the stark problems that AI presents to assessment in HE. The idea of verifiable judgement, if anything, signifies that assessors need to be certain that judgement has been developed and exercised by students/graduates. To put it more starkly, if we cannot guarantee the validity of assessment – and I suggest that particularly in the humanities that many universities would struggle to do this – then we are setting the sector up for unsolvable problems. Prof. Villena is right to call out the OfS and the QAA here. He might add university SLTs. I note that the HEPI blog on quality of a few days ago went without any comment on AI at all. Surely, in the QA realm, all other work must cease until we come to defensible solutions?
I tend to disagree with the conclusions that are drawn, generally and here, about forms of assessment. Prof. Villena artfully includes examinations in the “proxies” category, despite the fact that examinations, properly designed, could be placed in the “demonstrations” category. Perhaps one of the exam questions could be this: “If so much assessed work can be produced by a machine, the assessment was probably not measuring what we thought [discuss]”. Let us quickly note the constant usage of “measurement”, surely an uneasy bedfellow with the notion of “verifiable judgement”. The sentiment here is “if a machine can do it, it’s not worth doing”. But this is evidently not the case. There are conditions, which we had for a long time, under which inferences about the work and about the student were possible from their submissions – the products that they submit, if you like. Professional judgement allows us to infer the work that must have gone into this assessment for it to be as it is. (Certainly, this is expedited by discussing works-in-progress with students, which Prof. Villena includes in the “iterative” and work-by-portfolio.) What we are often asked to do now is mark the work that we see happening. Can learning be “seen”, live? Yes. There must be a place, though, for assessment forms that support slow learning, reflection, and most of all writing. Writing is something that is seemingly supported in this blog; however, despite the essay being a genre that is a “proxy”, this also seems to be true: “A student who produces a competent essay with AI assistance has not demonstrated that capacity.” As a proxy form of assessment, wouldn’t the words “with AI assistance” be redundant here?
A more minor point is the requirement for “explicit criteria” and “well-defined / transparent learning outcomes”. Why crowbar these in? As usual I invite anyone using the “explicit” or “transparent” shibboleth to explain what they mean by it in the circumstances they are discussing.
I am sympathetic to the critique made of modularity in this blog. Space no doubt prevented Prof. Villena from saying more about how this might be taken further. I have discussed ideas similar to that of insisting on “a minimum proportion of assessment that is demonstrably non-delegable” (see: https://pvswrites.wordpress.com/2026/03/16/a-provisional-solution-for-the-use-of-ai-in-university-assessment/). The two-lane approach adopted early on by the University of Sydney is another. The obvious antidote to assessing on a modular basis is to have programme-level assessment; this might involve under-assessing students in their taught courses, followed by assessment centres which use a variety of assessment types and evaluation by teams of academic, and other colleagues.
My only proviso concerning modularity is that replacing it would seem to make life-long learning more rather than less difficult.
Reply
Jonathan Alltimes says:
The argument promises better quality and more trusted qualifications, if assessment is redesigned to eliminate generative AI as a substitute for the capabilities of students. No one outside of a small number of subject-specific academics knows what the qualification means in terms of knowledge, as it relys on a shared cultural heritage, for which a degree is only a partial transmission mechanism.
The normal exemplars people have in mind when posting blogs about AI is the humanities and social research. Written assessments are pointing towards preparation for the written examinations, the written examinations originate in the requirements for marking lots of papers for lots of students, consistency in examination, moderation by the college of examiners, and variation when setting examinations. The basic capability is to write well: look at the gap between the standard of written assessments and learned texts. Higher education will use better detection and it will redesign. The answer offered is too general, assessments should match the constitution of the subject and be task specific. You can only defend against the uncritical use of AI, when you provide a working alternative for performing tasks with which the student can make comparisons for evaluating generative AI for what it is, how it works, and it differs from personal reasoning. Practically no student is formally taught argument and what counts as good subject-specific reasoning.
There is likely to be a little movement towards what is argued, but given the scale of the number of students, it would be too time consuming. What people know, can only be assessed by those who already know and to know is to do, actions must be causally reliable. The bureacratic rigmarole for the standardization of what is taught and learnt is a general accountability mechanism between providers and the OfS for the DfE and Parliament. It is not analogically equivalent to teaching and learning, it is like a set of general probes designed for a different purpose and probes themselves can be in error. Academics who are also practitioners are also subject to external verification, which includes research publications.
Who will pay for sustained education and lifelong learning in businesses?
Reply
Add comment