A university tells students that it wants them to understand artificial intelligence, but its policy may read like an instruction manual for avoiding punishment. That is the tension identified by researcher Sam Illingworth in a study of the rules UK higher education institutions publish about AI. The report, released by the Higher Education Policy Institute in May 2026, argues that many policies use the language of learning while placing their practical emphasis on disclosure, monitoring and disciplinary action. The question is not whether universities should prevent cheating. It is whether university AI policies are helping students develop the skills they will actually need. HEPI's account of the study provides its central findings and explains the author’s approach.
The work is an analysis and interpretation of institutional documents, not a direct measurement of student learning or the fairness of individual disciplinary cases. A policy that the researcher could not find through public search might still exist behind a university login. A policy that mentions enforcement may also contain valuable educational guidance. Those distinctions matter because the report’s sharp criticism should be weighed against the scope of the evidence it examined.
University AI policies should be understandable before an assignment starts
For a student beginning coursework, an accessible policy is not a luxury. They need to know what sort of help is permitted and how to describe it. If one module welcomes AI-generated explanations while another prohibits similar assistance, the difference should be explicit. When the only clear advice is a warning about investigation, students may become more anxious without becoming better informed.
Illingworth began with 163 UK institutions holding degree-awarding powers and searched for accessible AI policies on 16 February 2026. He located 96 for analysis. The report says 41 per cent of institutions lacked a publicly accessible policy under his search method. That does not establish that those institutions had no internal policy. It means a prospective student, parent or other outsider could not readily inspect the guidance through the approach used.
Accessibility matters because students often need to understand conditions before joining a course. Families and applicants may want to know whether a university supports responsible use or takes a particularly restrictive approach. Hidden documents leave them guessing. Equally, an institution might have legitimate reasons to keep internal guidance within a student portal, so the public-search test should not be confused with a legal ruling that an obligation has been breached.
A learning policy can still operate like a disciplinary rule
The researcher examined vocabulary across the 96 collected documents and read a smaller group closely. An automated count initially suggested a strong educational emphasis in many policies. Yet his closer reading found that educational phrases sometimes surrounded rules primarily concerned with recording use, proving authorship or detecting violations. The report argues that where a policy sits within a university’s administrative structure can reveal more about its purpose than the friendly words appearing in its introduction.
A document housed under academic misconduct will naturally be read in a particular context. Its function may be to define offences and protect assessment integrity. There is nothing inherently wrong with such a document. The problem arises if it is the only visible explanation students receive while an institution claims to be teaching critical understanding of AI. Enforcement and education are different tasks, and both need appropriately designed guidance.
The HEPI account says the researcher examined 77 keywords across the policies and reviewed 19 documents more closely. Those methodological details help readers understand why a large automated language count could not carry the whole conclusion. A word such as ‘learning’ can appear in a sentence that mainly tells students what evidence they must retain in case of investigation. Context changes meaning.
Trust is not the same as abandoning standards
Universities have legitimate reasons to protect the value of a qualification. Someone submitting work they cannot explain may not have met a course’s learning goals. Teachers need methods to assess original understanding, and students deserve to know that their own effort will not be devalued by unacknowledged shortcuts. A trust-based approach therefore does not require universities to accept every AI-assisted submission without question.
It asks instead whether rules begin by assuming students can act responsibly when expectations are clear. An assessment may permit AI for brainstorming but require the final argument and evidence to be independently produced. Another may ask students to compare a machine-generated answer with original sources, deliberately making verification part of the educational task. These choices can be demanding while treating learners as people acquiring judgement rather than potential offenders.
Illingworth’s report identifies institutions whose guidance he regards as more constructive examples, including the University of Stirling and Arts University Plymouth. Their inclusion is not proof that they produce the highest graduate outcomes or that their policies are beyond criticism. It shows different administrative and educational approaches are already possible within the sector.
An HEPI response by Open University academic Mirjam Hauck welcomes the central concern but argues that calls for ‘critical AI literacy’ need an operational framework. In other words, encouraging students to think critically is not enough unless universities explain what skills should be taught, practised and assessed. That response adds a useful perspective rather than treating the first report as the final word.
Different students experience the rules differently
A language model can be used to summarise instructions, reorganise a difficult passage, translate unfamiliar wording or practise the structure of an argument. Those uses may matter particularly to students who work in an additional language or need accessibility support. A policy that treats all computer assistance as equivalent risks blurring legitimate help with substitution for assessed work.
Fairness also means protecting students from unsupported allegations. Software that guesses whether text resembles AI writing cannot establish authorship with certainty. Universities should use reliable evidence and give students a meaningful opportunity to explain their process. An expectation to keep notes or drafts may be reasonable in some assessments, but administrators should consider whether the requirement creates unequal burdens for different learners.
LiveAIWire has examined the wider challenge of AI-related cheating rules and the tension between student empowerment and dependence on assistants. The policy study focuses on the institutional side of the relationship: not what students privately say about a tool, but what rules they are expected to follow and whether those rules help them learn.
The meaning of AI literacy must be practical
A student who is AI-literate should be able to recognise when a tool has produced an unsupported claim, explain why a source is credible and decide whether using the tool is appropriate for the task. They should know how to protect personal and confidential information. They also need to understand that an AI assistant can sound helpful while being wrong. These are judgement skills, not badges for completing an online demonstration.
Universities can teach them through assignments that require verification, comparison and explanation. For example, a student might be asked to identify weaknesses in a generated summary and defend a revised argument using original evidence. An employer could value that ability more than familiarity with a particular product name. Such assessment designs would also make permitted AI use easier to communicate because they focus on the intellectual work the student must demonstrate.
The challenge is to put those ambitions into departmental rules that a person can actually follow. Policy writers need to decide what forms of assistance are acceptable, how use should be disclosed and what to do when a case is ambiguous. They also need to ensure guidance can be found, understood and updated as technology changes. A beautifully phrased principle is little help if students cannot discover how it applies to tomorrow’s deadline.
The next test is what students can do afterwards
The HEPI study raises a defensible criticism about the structure and public accessibility of AI rules. It cannot tell us by itself whether particular universities are improving students’ long-term skills. That would require further evidence: interviews, analysis of actual assessments, and perhaps comparisons of what learners can demonstrate before and after structured teaching. A policy inventory is a starting point for accountability, not the complete measure of success.
The international policy debate provides another point of comparison. LiveAIWire has covered a separate education-policy effort to manage AI in schools, but school rules and university assessment standards have different purposes. A university normally expects students to develop independent disciplinary expertise while using specialist sources and methods. Its guidance therefore needs to explain not just what software is allowed, but what understanding students must personally demonstrate. Borrowing a generic prohibition from another sector may be administratively easy while doing little to make that distinction clear.
The more useful question for universities is whether their rules teach people to exercise sound judgement in an AI-rich world. Academic integrity remains essential. But if a student completes a degree knowing only how to avoid a disciplinary warning, they may still be poorly prepared to use the technology responsibly at work. Clear, accessible guidance that teaches verification and ownership of ideas would better serve both education and fairness.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.