May 25, 2026
What University AI Policies Are Really Teaching Students
A reflection on 96 UK University policies
Reading Time:
10 Minutes
Category:
Ai in Education
Policy is pedagogy
What University AI Policies Are Really Teaching Students
A recent HEPI Policy Note by Professor Sam Illingworth deserves the attention of every university leader. It studies 96 publicly accessible AI policies across UK higher education and asks not only what those policies say but also what they actually do. That distinction matters. Institutions often describe their AI guidance in the language of learning, support, responsibility, and skill. Yet when one reads closely, some of these policies function less like educational invitations and more like disciplinary systems.
As someone entrusted with institutional leadership, I find the study important because it shifts the discussion from tools to universities themselves. We have spent the last three years asking whether students may use ChatGPT, how they should declare AI assistance, and whether detectors can identify machine-generated text. Those are not trivial questions. But they are second-order questions. The first-order question is more demanding: what kind of learner does our policy imagine, and what kind of university does our policy reveal?
Illingworth's study began with 163 UK institutions with degree-awarding powers. On the day of collection, 96 had publicly accessible AI policies, while 67 did not have one that could be found through the study's public search method. That finding alone should trouble us. If AI is now part of academic life, then a policy hidden behind a login may serve current students, but it does little for applicants, parents, employers, regulators, partner institutions, or the public. Universities are public trust institutions. Our rules about knowledge production should be publicly intelligible.
The more striking finding is methodological. A computational analysis found that 86% of accessible policies appeared to be education-dominant by vocabulary. In other words, the most common words were learning, guidance, support, and skills. But a close qualitative reading of 19 policies complicated that result. Some policies that sounded educational at the surface operated through obligations, declarations, evidentiary burdens, and misconduct procedures. The study names this pattern performative education framing: the language of support, placed inside the machinery of control.
This is a subtle but consequential discovery. It indicates that a university can sound progressive while training students into defensive compliance. A policy can repeatedly use the word “learning” and still teach students that the central relationship between the learner and the institution is one of suspicion.
Policy is pedagogy
University leaders sometimes treat policy as administrative infrastructure. It is more than that. Policy is pedagogy in institutional form. It tells students what we value, what we fear, what we trust, and what kind of judgment we expect them to develop.
John Dewey argued that education is not preparation for life but a process of living, a continual reconstruction of experience. Paulo Freire warned against educational systems that reduce learners to objects of management rather than subjects of inquiry. Gert Biesta, more recently, has argued that education serves not only qualification and socialization, but also subjectification, the emergence of the student as a responsible subject capable of acting in and with the world.
These thinkers help us see why AI policy cannot be reduced to a list of permitted and prohibited acts. If our policies merely instruct students how to avoid accusation, then we are not educating them into responsible agency. We are teaching institutional risk management.
That does not mean universities should be naive. Academic integrity matters. Degrees must remain trustworthy. Students who submit work they did not meaningfully produce violate the relationship between effort, learning, and recognition. But integrity is not strengthened when the institution treats all students as latent offenders. The International Center for Academic Integrity defines academic integrity through values such as honesty, trust, fairness, respect, responsibility, and courage. It is revealing that trust appears in that list not as a sentimental add-on, but as a structural condition of moral community.
The hard task, then, is to build policy that protects standards without making suspicion the grammar of higher education.
The detector temptation
The attraction of AI detection is easy to understand. It promises administrative clarity at the precise moment when authorship has become harder to see. Yet the evidence base does not justify making detection the center of institutional response.
Dalalah and Dalalah found that false positives and false negatives are real possibilities in generative AI detection, with genuine and AI-generated text distributions showing meaningful overlap. Liang and colleagues found that GPT detectors were biased against non-native English writers, reporting an average false-positive rate of 61.3% on TOEFL essays in their tested sample. [6] Walters's comparison of 16 AI text detectors found that some commercial systems performed better than others, but many tools struggled, especially with GPT-4 and undergraduate essays.
The philosophical problem runs deeper than technical accuracy. Michel Foucault's analysis of examination and surveillance in Discipline and Punish remains useful here, not because universities are prisons, but because assessment can become a mechanism through which people internalize the gaze of authority. When students begin writing not to think clearly but to avoid seeming machine-like, the detector has already reshaped learning. The institutional danger is not only wrongful accusation. It is the production of a hidden curriculum in which students learn to perform humanness for automated suspicion.
This is especially serious for international students and multilingual writers. If detectors tend to penalize predictable prose, students still developing academic English may be at disproportionate risk. A university that says it values inclusion cannot build integrity systems that make linguistic differences look like misconduct.
The deeper issue is not cheating. It is the future of assessment
AI has exposed something universities already knew but often postponed: many assessments were built for a world in which producing text was a reliable proxy for thinking. That world has changed. The response cannot be nostalgia.
The most serious institutions will now distinguish between at least four types of academic tasks. Some tasks should be prohibited because they are designed to develop unaided fluency. Some should permit AI as a limited support, much as calculators or proofing tools are permitted under defined conditions. Some should require AI use because critical collaboration with tools is itself the learning outcome. Some should assess process, oral defense, design judgment, lab performance, placement practice, or reflective decision-making rather than the final artifact alone.
This is where the study's policy critique becomes strategically important. If we house AI guidance mainly within misconduct frameworks, we direct institutional energy toward detection, declaration, and punishment. If AI guidance is housed within teaching and learning architecture, it directs institutional energy toward assessment redesign, staff development, student voice, and critical literacy.
That distinction will define the next phase of higher education.
By the 2026-27 academic year, I expect serious universities to move from generic AI rules to assessment-level AI conditions. Every assignment brief must state whether AI is prohibited, permitted, encouraged, or required, and why. By 2027-28, I expect many programs to introduce at least one substantial assessment in each year where students must demonstrate how they used, challenged, rejected, or revised AI outputs. By 2028-29, AI detectors will likely be treated, where they remain in use, as weak signals that require contextual human judgment rather than as quasi-forensic evidence.
These projections are not speculative drama. They follow from the convergence of student behavior, detector unreliability, and policy lag. HEPI reports that 95% of UK full-time undergraduates already use AI in some form, and 94% use generative AI to help with assessed work. The practice is already here. The question is whether universities will make it educational.
Critical AI literacy, not tool compliance =
Long and Magerko define AI literacy as competencies that allow people to understand, use, and critically evaluate AI systems. UNESCO's guidance on generative AI similarly calls for human-centred policies that protect agency, equity, privacy, and pedagogical appropriateness. These are not decorative principles. They point to a different institutional posture.
Critical AI literacy asks students to understand what a model can and cannot know, how training data shape outputs, why hallucinations occur, how bias is reproduced, when automation narrows judgement, and what kinds of human responsibility cannot be outsourced. It also asks students to consider environmental costs, labour conditions, intellectual property, accessibility, and the political economy of educational technology.
That is far more demanding than telling students to declare AI use in a footnote. Declaration may be necessary, but it is not sufficient. A student can declare a tool and still be intellectually passive. Another student can use a tool critically, minimally, and responsibly, while learning more about epistemic humility than a prohibition-only policy would ever teach.
The issue is not whether students used AI. The issue is whether they remained authors of judgment.
Hannah Arendt described thinking as an inner dialogue that enables responsibility in the world. In universities, we should be wary of any practice that interrupts that dialogue, whether the interruption comes from a chatbot that supplies premature fluency or from a policy regime that turns reflection into compliance documentation. Both can weaken judgment. The purpose of policy is to make better judgments more likely.
What should presidents and senior teams do now?
For university leaders, the study offers a practical challenge. We should audit our AI policies not only for content, but for institutional function.
Leadership question | Why it matters |
Is the policy publicly accessible? | Public trust requires visible rules about knowledge, authorship, and assessment. |
Is it located primarily in misconduct procedures or teaching and learning strategy? | Location shapes interpretation. A supportive sentence inside a disciplinary framework may still function as control. |
Does the policy begin from trust or suspicion? | Students learn the moral atmosphere of the institution from procedural design. |
Does it include student voice? | Students are not merely policy subjects. They are participants in the emerging culture of AI-mediated learning. |
Does it define AI literacy beyond tool operation? | Tool training without critical understanding produces competent dependence, not educated agency. |
Does it redesign assessment rather than simply police output? | AI has weakened output-only evidence. Assessment must recover process, judgement, and context. |
The most credible AI policies will not be the longest or the most restrictive. They will be the clearest about educational purpose. They will tell students where the boundary lies, but they will also explain why the boundary matters. They will distinguish cheating from poor judgement, poor judgement from experimentation, and experimentation from serious learning. They will protect academic standards while refusing to make mistrust the default relation between student and university.
That is a demanding balance. It is also the work of educational leadership.
The university after the AI panic
The first stage of institutional AI response was panic. The second was prohibition. The third was procedural compliance. The next stage must be educational reconstruction.
A university worthy of the name cannot define its future by asking how to catch students using the dominant knowledge technology of their time. It must ask how students can learn to think with and against that technology, how they can preserve agency amid automation, and how institutions can assess learning without reducing education to textual output.
The HEPI study is useful because it holds up a mirror. Many universities may find that their policies say “learning” while their structures say “surveillance.” That contradiction is not a communications problem. It is a governance problem, a pedagogical problem, and ultimately a moral problem.
Our students are entering a world in which AI will shape writing, coding, research, design, administration, medicine, law, finance, and public culture. They need more than warnings. They need formation. They need the intellectual habits to ask what a machine has optimized, what it has omitted, whose interests it serves, and when a human being must refuse its convenience.
If our policies teach that, they will be worthy of the university. If they teach only how not to get caught, then the failure will not be the students'. It will be ours.
References
[2] Sam Illingworth, UK University AI Policies: Open Data and Analysis, GitHub repository
[3] Professor Sam Illingworth, Edinburgh Napier Research Repository
[4] HEPI and Kortext, Student Generative AI Survey 2026
[6]: https://www.cell.com/patterns/fulltext/S2666-3899(23 )00130-7 "Weixin Liang, Mert Yuksekgonul, Yining Mao, Eric Wu, and James Zou, GPT detectors are biased against non-native English writers, Patterns, 2023"
[7] UNESCO, Guidance for generative AI in education and research, 2023
[9] International Center for Academic Integrity, The Fundamental Values of Academic Integrity
[10] John Dewey, Democracy and Education, 1916
[11] Paulo Freire, Pedagogy of the Oppressed
[13] Michel Foucault, Discipline and Punish: The Birth of the Prison
