How are students using Generative AI in UK universities?

Honestly I’m not sure how worried we should be about these findings from HEPI (n=1,041) given it seems the sector has got passed its initial inclination to try and prohibit. If we’re in a situation where only 12% of students are not using LLMs in their assessment then what matters is steering use towards epistemic agency* and way from LLMs supporting a turbo-charged transactional engagement with knowledge.

It’s interesting to contrast these findings with Anthropic’s study of university students using Claude, classified in terms of Bloom’s taxonomy:

The dynamics of cognitive outsourcing (and potential lock-in) differ as you move up from lower to higher-order thinking skills for students. I struggle to see a problem with students using LLMs to support understanding materials, much as I struggle to see a problem with academics using LLMs to produce materials which are easier to understand. Sure we might rapidly end up in a situation where this learning interaction is mediated by LLMs by default but I don’t see a fundamental difference in type from that being mediated by other kinds of digital platforms (e.g. the LMS) or outputs (e.g. Powerpoint). It’s a case of better or worse design rather than something human being lost through the introduction of a technological element.

I think applying and analysing by definition lend themselves to agentive engagements with knowledge. You can’t get the LLM to do something useful unless you’re thinking about what you’re asking, which means to at least some extent an epistemic capacity is being exercised. Certainly students could try and fail to do this, but that’s a different kind of problem to be addressed through the register of AI literacy. The pedagogical challenge comes in recognising how students are doing this in order to design learning processes which support increasingly purposive applications rather than just assuming they will be learning in the same way we did.

It’s evaluating and creating where it gets more concerning. If you’ve already developed these capabilities LLMs can be used to speed up the process (though a soft lock-in might result over time) or enhance the process in the activity I describe as rubber ducking. The problem arises if you haven’t learned how to do this without the LLM, such that the composite capacity (e.g. writing a report) develops in a way that has the LLM baked into it from the outset. For example reliance on LLMs for an outline only concerns me if students haven’t learned to do this without the LLM in the first place. To rely on it to critically evaluate your work and suggest room for improvement carries a similar risk of cognitive outsourcing which is unlikely to be addressed after university by most students.

This is a long-winded way of saying that we urgently need to get beyond the category of ‘AI’ in how we think about these pedagogical challenges. The relationality within the LLM becomes more important to recognise the further up the taxonomy we go. Exactly what ‘creating’ means can now vary immensely depending on the pattern of interaction the student has with the LLM.

It’s also interesting to see that:

• “The main factors putting students off using AI are being accused of cheating (said by 53% of respondents) and getting false results or ‘hallucinations’ (51%). Just 15% are put off by the environmental impact of AI tools.
• Students still generally believe their institutions have responded effectively to concerns over academic integrity, with 80% saying their institution’s policy is ‘clear’ and three-quarters (76%) saying their institution would spot the use of AI in assessments
• The proportion saying university staff are ‘well-equipped’ to work with AI has jumped from 18% in 2024 to 42% in 2025.

I think students are over-estimating how effectively institutions can identify (and act!) on problematic LLM use and over-estimating the AI literacy of academic staff. If I’m right and student perception catches up to that reality, could ‘cheating’ as an inhibiting factor start to collapse from that figure of 51%?

*Thanks to my collaborator Peter Kahn for introducing me to this notion