A university student may use artificial intelligence to understand a difficult reading, plan an essay or prepare for an examination, all before attending a single lecture. That is no longer unusual behaviour. A 2026 survey from the Higher Education Policy Institute, or HEPI, finds that generative AI has become a normal part of studying for the vast majority of UK undergraduates it questioned. But its findings also reveal a striking gap between what students already do and the guidance they believe they receive. AI at university is becoming ordinary before institutions have settled what responsible use should look like. The HEPI Student Generative AI Survey 2026 offers a valuable, if limited, picture of that change.
The survey is based on responses from 1,054 full-time UK undergraduates collected by Savanta in December 2025. It describes reported behaviour and opinion within that student sample. It cannot independently establish how much knowledge individual students acquired, whether their coursework met a particular academic standard or how every university in the country applies its rules.
AI at university is often used as a study companion
Ninety-five per cent of respondents said they had used AI in at least one way, while 94 per cent reported using generative AI to help with assessed work. Those figures do not mean that nearly every assignment was written by a chatbot. Help can range from explaining a concept to proposing an outline or preparing practice questions. The important distinction is between using software to support one’s understanding and submitting machine-generated material as though it were personally produced work.
That distinction is not always easy to draw. A student who asks for feedback on a paragraph may learn why their argument is weak. Another who asks a model to produce the final paragraph and copies it may have bypassed the learning the assignment was meant to develop. The same technology enables both patterns. That is why a simple institutional rule of ‘AI allowed’ or ‘AI forbidden’ can be too crude to describe how an activity should be assessed.
The HEPI survey reports that 12 per cent said they directly included AI-generated text in assessed work, compared with eight per cent in the 2025 survey and three per cent in 2024. This trend concerns a self-reported behaviour across separate survey waves, not a direct examination of every submitted assignment. It shows why universities are under pressure to explain what acceptable assistance looks like.
Students value the help and worry about dependence
Almost half of surveyed students believed AI improved their university experience. They described convenience, faster access to explanation and support in making sense of challenging material. Other respondents raised concerns about reduced effort, fairness, isolation or the possibility of losing skills they would need after graduation. The contrast matters: one person’s efficient learning aid can become another person’s shortcut around learning.
HEPI reports that 68 per cent considered AI skills essential for thriving in the modern world, while fewer than half said teaching staff were helping them develop those skills for their future careers. The numbers are about student perceptions. They do not establish that every lecturer lacks knowledge or that university staff have failed to provide any training. They do suggest that a significant share of students cannot see a clear bridge from today’s tools to tomorrow’s professional expectations.
A robust education must include being able to question an AI answer. Students need to check sources, recognise unsupported claims and explain decisions in their own words. These habits are important whether the subject is history, law, engineering or business. Merely showing someone how to enter a clever prompt does not teach them how to detect an invented citation or a misleading statistic.
The previous HEPI survey published in 2025 recorded a rapid rise in student use for assessment-related tasks. Reading the two reports together makes the pace of change easier to understand. The later results show that the debate is moving from whether students use AI towards whether universities have given them the intellectual and practical tools to use it well.
Clear rules can protect learning and fairness
A student should be able to find an assessment’s AI policy before starting the work. The guidance should explain whether brainstorming, translation, grammar support, coding assistance or generated prose is permitted, and what must be acknowledged. Policies that rely on vague phrases leave room for inconsistent interpretations. A rule that is clear in one module but different in another may still be justified, provided the reason is explained.
For students with accessibility needs, AI support may have a different significance from its use as a simple convenience. Institutions need to distinguish legitimate assistance from dishonest substitution without making students disclose irrelevant private details. The aim should be fair assessment of learning outcomes, not a competition over which student can use an approved tool most discreetly.
Worries about false accusations also deserve attention. Automated AI detectors are not reliable judges of authorship in every context. Disciplinary procedures need evidence and proportionate human review rather than an assumption that a software score proves misconduct. This is a matter of procedural fairness, irrespective of whether an institution favours extensive or limited AI assistance.
LiveAIWire has previously examined students' simultaneous sense of empowerment and dependence on AI, along with concerns over how schools respond to AI-assisted cheating. The HEPI findings add another layer: students are not merely choosing between honesty and convenience; they are navigating rules, expectations and future skills that may not yet be consistently taught.
The unexpected question about loneliness
The survey also asked about students’ experience of AI and companionship. Around 15 per cent reported using AI for companionship, advice or to address loneliness. Responses on whether AI made people feel more or less lonely were split, with one-fifth saying it increased loneliness and a similar proportion reporting the opposite. These are subjective survey answers, not evidence that the technology caused either outcome for every respondent.
That should discourage easy conclusions. A chatbot may provide a useful way to practise a conversation or organise difficult thoughts. It cannot automatically replace friendship, university support or professional help. Researchers would need longitudinal evidence and careful comparison groups to understand whether use changes social relationships over time. HEPI itself calls for more work on the issue rather than announcing a settled benefit or harm.
The broader educational question is what universities want students to become. If higher education is meant to build independent judgement, the ability to interrogate software may be a core skill. That does not mean students must reject tools, any more than they must reject calculators or digital libraries. It means the course should make clear what knowledge a graduate needs to own and what assistance can sensibly be delegated.
Universities need to teach decisions, not just tools
A practical approach could give students examples of permitted and prohibited assistance for each kind of assessment, together with opportunities to practise verification. Staff need time to learn the tools and understand their limits. Students need access that does not depend solely on their ability to purchase expensive subscriptions. Institutions should also ask whether their assignments still test the abilities they intend to develop.
Students should be expected to explain significant claims, identify sources and recognise when machine assistance has changed their reasoning. Those expectations can be demanding without being punitive. They give learners a more durable skill than familiarity with any individual product, whose interface and capabilities may change long before graduation.
That challenge is also about access. Not every student can afford an expensive subscription, and some may rely on a university-provided account with different restrictions. If teaching assumes that everyone has the same tools, the institution may unintentionally build another inequality into coursework. Equitable policy would identify which uses are essential, provide appropriate access where possible and ensure assessments remain fair for learners who choose not to use a particular commercial service.
It would also help to link this discussion to the changing shape of university qualifications. Employers may increasingly ask graduates to demonstrate that they can challenge a generated answer, trace evidence and explain how they reached a conclusion. Institutions cannot assume that the mere presence of AI on a campus develops those skills. They need opportunities for practice, transparent marking criteria and examples showing the difference between legitimate assistance and handing over the substance of a task. These are educational proposals arising from the survey, not outcomes proved by it.
HEPI’s findings describe a system already living with AI while still negotiating how to educate for it. The most useful response is neither blanket enthusiasm nor blanket prohibition. It is clarity about what students are learning, honest support for good uses and firm standards for work that must remain their own.
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.