By Miles Thompson
We know that debates around AI and the university are here to stay. In a much more zoomed in conversation, this blog centres around an interesting experience I had with “AI overviews”. This experience made me think about their use and potential misuse more generally, especially how busy students might respond to them. This blog also highlights: i. a recent Radio 4 programme on AI and UK universities and ii. a new report on the environmental sustainability of AI in post-16 education.
The Artificial University
Beginning with the broader issue, earlier this month, Radio 4 broadcast a programme called “The Artificial University”, presented by Dr Kathryn Claire Higgins (Goldsmiths). She spoke about the challenges of balancing AI and the purposes of the university. Within this large debate, one thread really struck me.
Today’s students were described as juggling other responsibilities, with worries about debts, future jobs and earnings. It echoes research I have published about students generally (link 1, link 2). Against this backdrop, AI was seen as a tool students might use either willingly or because they worry they will be left behind if they aren’t using it. There was a fear, from some, of using AI incorrectly, and of being accused of academic misconduct. As a result, some students now build up portfolios of proof that their work is done by them and not AI.
The programme was a really interesting listen and is available on BBC Sounds. On a parallel but different track, this blog focuses in on how AI overviews might sometimes close down rather than open up avenues of enquiry for students and others.
When AI tried to summarise my research
In short, I was trying to confirm whether a doi (digital object identifier) was working correctly on a recently published output (answer: it wasn’t, but is now). During the process, I put the full title of the article, in quotations marks, into google (“A conversation with ourselves: reflecting on our professions, the systems we work in and our roles within it”). The article is currently paywalled on the publisher’s website (here), but the author accepted manuscript can be accessed by anyone, including AI, through my institution’s workplace repository here. At the top of the search results, an AI overview presented what appeared to be a summary of the work I wrote with my co-author, and I was taken aback by what I saw.
The article is a reflective piece, aimed at clinical and counselling psychologists working in the NHS and/or people seeking to do the same in the future. The piece attempts to contextualise and present some statements reflecting critical debates, hopefully providing the basis for further discussions. The statements are:
- 1. The NHS can perpetuate inequality.
- 2. We draw power and security from this system.
- 3. Limited / limiting questioning.
- 4. The mental health diagnostic system lacks scientific validity and causes harm.
- 5. The difference that doesn’t make a difference.
- 6. None of this is new.
- 7. Working today.
- 8. Daring to organise.
As you can see, the piece is quite NHS and therapy specific, but here is what AI said about the article:
The first paragraph is accurate. But the summary of article: 1. Reflecting on your profession; 2. Understanding the system; 3. Your role in the system – barely reflects the article content. And yet at the end, the AI links to the UWE research repository where the full article can be downloaded.
Surprised by the result, I refreshed the browser – inputting the search term again. This time:
Again, it starts with some accurate information about the article title and authors – before providing 3 summary ideas: 1. Map your system; 2. Spot the rules; 3. Change your role. Which has little resemblance to the 8 statements noted above.
Conscious of AI’s impact on the environment (more soon), I repeated the search one final time:
The same pattern. Some general accuracy at the start, but this time the main points of the article are apparently: 1. The Job; 2. The system; 3. Your role.
We know AI gets things wrong. Indeed, it says at the bottom of each of the images: “AI responses may include mistakes”. What struck me here was that I wasn’t asking AI for a summary of the article. Actually, I wasn’t asking for AI for anything at all. And yet an AI overview was provided, generating what appeared to be a summary of the article. But, it was not the contents of the article my co-author and I wrote.
Stepping back a bit, the AI overview likely reflects the model’s prior training rather than the content of the actual opinion piece. Generally speaking, AI overviews generate responses by combining information from multiple sources together with patterns learned during LLM (large language model) training. In this way, the overview was probably summarising material that is contained in general work-based-reflective-pieces – rather than the specific material the actual article contained. But this is not how the overview positioned itself. Arguably it presents itself as a summary of the actual piece. It also provides links to where the article can be downloaded – perhaps reinforcing the idea that it had retrieved information directly from the source.
As one of the authors, of course I knew the summary was wrong – and I found its incorrectness interesting. But, having listened to the Radio 4 programme above, I also wondered what I would have thought of it if I was a time poor student, trying to hunt down relevant information for an impending deadline. Would this summary make it more, or less likely I took the time to download and look at the original article for myself? Less likely, I concluded.
The environmental dimension
Returning to the report on “AI and environmental sustainability in post-16 education”. It usefully and concisely highlights the environmental footprint of AI. From rare earth extraction to data centre building and cooling, to end of life e-waste. It talks about what institutions can do, and has a great section on possible personal actions (section 5). It notes how “not every task needs AI” and that “alternatives are often quicker, more reliable, and less resource-intensive” (Bonner et al., p.19). Concretely, the report highlights how easy it is to stop AI overviews. Simply by adding “-AI” to your search query, or by adding browser extensions that do this automatically, they will not appear. When I do this for the query I described earlier, I simply get links to the webpages where, with a bit of scrolling and a click or two – I can download the article and its actual statements.
Thinking about this episode again from the perspective of a time poor student, while accurate overviews can save us time and further searches, inaccurate overviews could shut down potentially useful avenues of enquiry.
[Overviews on] – I think a source might be of interest. I search for it. Get an inaccurate AI overview. Decide the source is not of interest. Move onto the next source.
[Overviews off] – I think a source might be of interest. I search for it. Get the links to it. Download it. Decide for myself whether it is of interest or not.
To repeat, students or other users who are led by an inaccurate AI overview may think they are seeing a summary of the core arguments of a specific article, when they might be seeing a summary of very general patterns that are present across multiple historical pieces. And yet, the specific article, and its actual information, is often just a couple of mouse clicks away. But does the user click or not? Do they make up their own mind or not? Maybe this blog, and others like it, will make it a little more likely that we wrestle back a small element of user discretion.
For me, I’m grateful for the additional information I have on the impact of AI on both my sector (Higgins, 2026) and the environment (Bonner et al., 2026). As well as the simple steps I can take to reduce my unnecessary use of AI. Remember, this blog is not about using or not using AI per se. Although its environmental impact should give us all pause, and caution, AI can still be useful. Instead, this blog seeks to raise awareness of when an AI overview may unintentionally discourage us from consulting the original source.
Finally, if anyone is interested in the actual article and reading the statements for themselves – the link is below (Thompson & Daffin, 2026).
References.
Higgins, K, C. (2026, July 6). The Artificial University [Radio broadcast]. Radio 4. https://www.bbc.co.uk/programmes/m002ykn6
Bonner, C. et al. (2026). AI and environmental sustainability in post-16 education. Jisc / EAUC. https://www.jisc.ac.uk/reports/ai-and-environmental-sustainability-in-post-16-education
Thompson, M., & Daffin, J. (2026). A conversation with ourselves: Reflecting on our professions, the systems we work in and our roles within it. Clinical Psychology Forum, 397, 4-10. https://doi.org/10.53841/bpscpf.2026.1.397.4 / https://uwe-repository.worktribe.com/OutputFile/16110094