This page has been migrated from a previous version of our website. If you spot any issues or can’t find what you were looking for by searching our site please contact us to create a Marketing Customer Service enquiry and we will be happy to help.
Professor Rose Luckin looks at the generative AI results from the first national pilot of a Pre-arrival Academic Questionnaire, funded by the Office for Students and led by the University of East London, in conjunction with Advance HE and JISC.
What incoming students actually know about AI: the evidence universities are not using
Ask most senior leaders in higher education what they know about their incoming students and generative AI, and you will hear one of two answers. Either: “they are all using it, and we have to manage that.” Or: “the situation is changing so fast that we have no real data.” Both answers, I would suggest, are in their different ways excuses for not looking at the evidence that already exists.
The National Pre-arrival Academic Questionnaire (PAQ) pilot, led by Dr Michelle Morgan at the University of East London, Advance HE and JISC, and funded by the Office for Students, is one of the most significant data-gathering exercises in recent UK higher education. In Wave 1, completed by over 5,500 incoming undergraduates across 15 English institutions in September 2025, it asked students directly about their generative AI experience before they had even started university. The findings could usefully sit on the desk of everyone responsible for AI strategy, curriculum design, access and participation, and student transition.
Most students have used generative AI but “most” conceals more than it reveals
Roughly three in five incoming undergraduates, 60.6 per cent to be precise, report having used generative AI in some form before arriving at university. That is a clear majority, and it has rightly attracted attention. But the more important number is the one sitting behind it: 39 per cent of incoming undergraduates have no experience of generative AI whatsoever. That is not a niche group. That is nearly two in every five students walking through your door.
In a sector that has spent two years debating how to respond to AI, we have not yet adequately reckoned with the students who are arriving with no foundation at all. If institutional AI policy, induction design and assessment reform are all calibrated around a student body assumed to be broadly AI-literate, then we are compounding disadvantage for a very large group from day one.
What students are actually doing with generative AI
Among those who do use it, the picture is notably less alarming than the dominant narrative suggests. The most common uses are learning-oriented: exploring topics of interest (51.6 per cent), correcting grammar and spelling (50.9 per cent), summarising information (43.9 per cent) and having concepts explained (42.9 per cent). These are study support functions. Students are using AI as a tutor, a proofreader and a reference tool, not primarily as a ghostwriter.
Generating drafts or rewriting work, the uses that attract most anxiety from academic integrity teams, account for around 15 to 16 per cent of use cases. Getting feedback on writing sits at 32.6 per cent. More technical uses, such as creating code or generating presentations, remain genuinely uncommon at 4 per cent and 3.8 per cent respectively, which makes sense for students who have not yet encountered the technical workflows where such tools are most embedded.
A cluster analysis of the data identifies four distinct user groups. Nearly half of AI-using students (48 per cent) are light users: they have experimented with one or two applications but AI is not embedded in their habits. A further 31 per cent are learning-focused users who use AI for topic exploration, summarisation and explanation but do not use it to generate or rewrite their own work. Writing support users (18 per cent) are a distinct cohort in which every student uses AI for drafting or rewriting. And at the top, a small group of power users (4 per cent) have already integrated generative AI broadly and deeply into their workflows.
The writing support users deserve particular attention from institutions designing assessments and academic integrity frameworks. They are arriving already habituated to using AI as part of the writing process itself. Policies and induction programmes that treat this as a future risk, rather than a current reality, are already behind. Crucially, recent research on students already inside the system suggests that whether those existing habits deepen or are redirected, depends heavily on what the assessment environment rewards.
The Trained to Stop Learning: How Students Are Experiencing Assessment and Learning in an Age of AI report published by Wonkhe in March 2026 evidenced that students who feel their course primarily rewards outputs, use AI at nearly twice the rate of those who feel thinking and reasoning are genuinely valued. The habits students carry through the door are not fixed. But they are shaped from day one by the signals institutions send.
ChatGPT and the problem of the single-tool mental model
Among PAQ students who use generative AI, ChatGPT is near-universal: 97.4 per cent have at least some familiarity with it, and over 82 per cent report meaningful experience. Nothing else comes close. Grammarly is a distant second at 58.7 per cent, followed by Google Gemini (44.4 per cent), Microsoft Copilot (41.4 per cent) and Canva AI (37.4 per cent). Claude, Midjourney and DALL-E all sit below 15 per cent.
This market concentration matters educationally. When students think “AI”, they are almost entirely thinking “ChatGPT”. That single-tool mental model shapes what they believe AI can and cannot do, what questions they know to ask of it, and how critically they evaluate its outputs. A ChatGPT-only understanding of generative AI is an impoverished one, and universities that simply accept it as the default starting point are not doing their students any favours.
There is a real opportunity here. The near-universal familiarity with one tool provides a shared reference point that course teams can use to broaden understanding, rather than starting from scratch. The question is whether institutions will seize it.
Adoption is not evenly distributed, and that should worry us
The PAQ Subject area is the strongest demographic predictor of generative AI adoption in this dataset. Students entering Computing (75.3 per cent), Mathematical Sciences (77.8 per cent), Physical Sciences (72.9 per cent), Engineering and Technology (70.0 per cent) and Business and Management (67.7 per cent) all show adoption rates well above the 60.6% average. Students entering Creative Arts and Design (46.4 per cent), Humanities and Liberal Arts (41.6 per cent) and Language and Area Studies (36.8 per cent) show rates significantly below it. Gender shows a meaningful, if moderate, association (Cramér’s V of 0.14), with male students reporting higher usage. Ethnicity also shows a small association, which may partly reflect differential access to technology or the particular profile of international students in the sample.
These are not dramatic differences, and it would be a mistake to overstate them. But they accumulate. A student from a less affluent background, studying humanities, who is the first in their family to attend university, faces a compounding set of factors that the aggregate adoption figure of 60.6 per cent entirely obscures. We do not yet have the full cross-tabulated demographic picture from the individual-level data, but the directional signals are clear enough to act on.
The 39 per cent are not an afterthought
Higher education has a habit of designing AI programmes, policies and frameworks around the students who are already most comfortable with technology. The PAQ data is a reminder that this approach risks leaving a very large group further behind.
The risk is not merely one of unfamiliarity with tools. Students who arrive without AI experience and then encounter institutional guidance that is contradictory, absent, or pitched at those who already know the landscape are precisely the group most likely to be penalised by that incoherence. The costs of unclear AI policy, as the emerging evidence published by Wonkhe on students already in the system makes plain, fall hardest on the most conscientious: those who try hardest to comply and receive the least support in doing so.
Basic orientation to generative AI tools, appropriate to disciplinary context and without assuming prior familiarity, could usefully be a standard component of induction. Not an optional extra. Not a bolt-on for students who ask. A standard component.
Three things universities could usefully do differently
First, universities could stop treating AI literacy as a problem to be managed and start treating it as a capability to be built. Build it from the point of arrival, with explicit attention to the students who arrive with least. The 39 per cent need an on-ramp, not a policy statement.
Second, universities could move beyond the ChatGPT monoculture in their own communications and teaching. Students who know only one tool, and know it uncritically, are not AI-literate. They are tool-dependent. Helping them understand a broader landscape, including the limitations and failure modes of the tools they already use, is part of the curriculum, not an addition to it. And it matters for learning outcomes as well as digital literacy: the evidence is becoming clear that students in environments which reward thinking over production engage with AI in fundamentally different ways, using it to interrogate and understand rather than to generate and submit. Building critical AI understanding from the start is part of creating those environments.
Third, universities could act on pre-arrival data. The PAQ exists precisely to give institutions intelligence about what their incoming cohort brings with them. GenAI readiness is now a meaningful dimension of that picture, and institutions that are not acting on it are passing up one of the few genuinely evidence-based opportunities to get ahead of the curve.
Wave 2 of the PAQ will follow in September 2026. The longitudinal comparison will be revealing. But the data from Wave 1 is already rich enough to act on. The question is no longer whether students are using generative AI. It is which students are being left out of that story, and what we intend to do about it.
Professor Rose Luckin is Professor of Learner Centred Design at the UCL Institute of Education and a member of the National PAQ Pilot Working Group.
