Search site...

AI is forcing us to rethink higher education: let’s do that

Will Stronge

10th September 2026

Introduction

Britain’s universities are in an economic crisis. The Office for Students reported in November 2025 that 124 providers, around 45% of the sector, expect to post a deficit in 2025-26, and that nearly one in six hold less than thirty days of liquidity.[1] Almost two in five institutions are now cutting staff, with more than 12,000 jobs lost during 2025.[2] The causes seem to be structural, with the tuition fee having been limited to £9,250 since 2017, losing roughly a quarter of its real value to inflation, plus a fall in international recruitment.[3]

A second shock has landed on top of this. By 2025, 88% of UK undergraduates reported using generative AI in their assessments, up from 53% a year earlier (and we can assume the numbers recede towards 0 if you go pre-2023/2022).[4] These two crises have created a complete mess of business as usual. The funding squeeze has exposed how much of the university depends on effectively selling a single product: the qualification (as devalued as it might seem today). And AI is now calling into question the value of this product, both in terms of what it ‘gets you’ once you’re in the labour market, but also in terms of what it takes to achieve it during the process of education itself.[5]

The practice of being a student has been fundamentally disrupted.[6] A student can now produce a competent essay in minutes, without having to have put much thought into it whatsoever. When academics come to mark it, some can identify the tell-tale signs of ChatGPT (some cannot), and in any case any complaints about the student’s work process are buried by the 40 other essays following close behind which are composed of somewhere between 5% and 90% of AI’s handiwork.

Cognitive decline?

This practical crisis within the university is sometimes linked to a perceived crisis concerning the effects of AI on our minds. Many are pointing to a recent paper which talks about the ‘cognitive atrophy’ that AI brings about due to the automation of a number of the (usually brain-led) steps of argumentation and articulation. The paper demonstrates that those who use AI first to complete a specific cognitive task have lower brain activity than those who are forced to go about the task without the support of AI – showing that we are cognitively offloading when we use AI rather than developing our internal capacity. One does not need to do the hard work of thinking, researching and writing to achieve whatever cognitive, reasoning task you have in front of you.

It should be noted, however, that within the tested groups of this study, the group with the highest levels of brain activity were those who employed both a period of critical reflection and a period of AI prompting (In comparison to groups which either used no AI or who jumped straight into AI to complete the task).

Worries about cognitive atrophy due to platform tech is not new: a very similar study was published in 2011, looking at the potential cognitive decline that might arise from the existence of the internet’s trove of information, accessed through Google search. We’ve had an easily accessible store of answers and knowledge for some time now: was the World Wide Web, in fact, the start of our cognitive downfall? I will return to this later.

Possible responses to these real and perceived crises

Nils Gilman has written an excellent article about American universities – specifically on their long-term, and more recently accelerated, decline.[7] He argues that the common responses to AI within higher education miss the point. One is defensive: ban the tools, return to invigilated exam halls (basically treat AI as cheating to be policed). The other, as Gilman represents it, retreats into the “student experience,” selling residential and campus life as the thing AI cannot replicate. 

The first response basically pretends that AI doesn’t exist outside of the university and will try to create policy responses (e.g. bans) that simply won’t work due to the ubiquity of these tools beyond the university’s jurisdiction; not only will students be able to circumvent anti-AI rules with some decent editing skills but also, and more importantly, students will be graduating into a world where some form of AI will be present (and therefore they will be expected to be somewhat AI-literate). The second “retreat” response dodges the question of the content, method and evaluation of learning itself, which is what students and (where there is public funding) the taxpayer ultimately pay for.

Responding to the wider worries about cognitive decline, I would argue that the defensive reactions to AI also often rest on a misconception about the mind. We have never thought using simply our brain and body alone. The mind reaches out and works through its tools: the pen and notepad, the calculator, the map, the google search and now the language model.[8] We are, in Andy Clark’s phrase, “natural-born cyborgs”: the brain is the core of an extended, collaborative system, not the whole of it (this is called Extended Mind Theory or EMT).

The flipside to this is that if you strip away the brain from the tools that it regularly latches onto then our minds are impoverished. A large language model is simply the newest and most powerful tool we think with and through in this sense (some more regularly than others of course). That fact doesn’t mean we should blindly adopt the technology – every new technology (particularly a general purpose technology) has more or less drastic effects on our minds, as figures such as Marshall McLuhan didn’t stop telling us. Nor does it mean that we’ve figured out how to use it best; right now most people will be treating AI as some kind of oracle that you pose questions to, rather than a function of a wider chain of thought. Nor should we ignore questions of ownership and control. I have argued that these questions become even more pressing in the age of monopoly platforms: large language models should be commonly owned, publicly backed and not based on a business model that demands huge natural resources.[9]

In short, I think the bigger, more difficult question is not whether students should integrate AI, but in what ways and to what ends they should do so.[10] When universities – or various commentators – basically demand a ban on AI outright they not only display a very narrow understanding of human education, but also of human cognition.

One final note before I discuss what should be done: even if the business models of Big Tech in Silicon Valley and in the Chinese tech elite should fail, we are still going to be left with this technology itself in some form. There are already many open source, low carbon-intensity models that are able to write the kind of essays that would pass examinations in universities up until now. These will continue to exist even if we don’t see further improved frontier models or the AI bubble bursts: we need to figure out how to reconfigure education in light of this new widespread capability.

The necessary return of radical pedagogy

With this in mind, the future of the university needs to reconsider some of the radical pedagogy of the past. In the 1960s Paulo Freire attacked what he called the “banking” model of education: the teacher deposits information into passive students, who store it and hand it back on demand.[11] Freire prefigured adjacent arguments made later by Sir Ken Robinson, who understood our model of education as one replicating industrial production (yearly batches of students, separate faculties that focus on different subjects, standardised tests etc.).

The marked essay and exam question are the banking/industrial model’s examples par excellence, and this is why AI now has shown them up as the mere set of standardised patterns that we currently ask students to recognise and reproduce.

Freire’s alternative was problem-posing education, in which teacher and students work as co-investigators through dialogue. The task here is not to receive answers but to pose better and better questions: I think this is a particularly valid approach to education (one I have integrated into my own teaching in the past) once answers are abundant and cheap (as in a world with AI capabilities). One thing AI cannot hand over is knowing what to ask of it and which directions of inquiry are valuable (although, of course, it can be an assistant to help you figure out these things).

Of course, education is also about the ability to retain and reuse what we learn – it’s not just about asking questions – but with much greater access to information within and outside the university, a further shift is likely necessary from ‘storage’ to ‘critical’ or ‘dynamic’ functions. Detouring back into the philosophy of mind and cognitive extension, we can think about what AI and its internet of sources mean to us as cognitive agents. Against the grain of mainstream scaremongering on the issue at the time, the authors of the 2011 “ur-paper”, if you will, on digital ‘cognitive atrophy’ wrote that:

“This is preliminary evidence that when people expect information to remain continuously available (such as we expect with Internet access), we are more likely to remember where to find it than we are to remember the details of the item. One could argue that this is an adaptive use of memory—to include the computer and online search engines as an external memory system that can be accessed at will.”

Without explicitly referring to the subfield, the authors are referring precisely to an ‘extended’ picture of mind (EMT), very similar to the one I have been utilising. It makes sense that we use our brain more to remember where to access information than to store that information itself: it’s a way of cognitively offloading huge amounts of information processing, as we have been doing for all of human existence with other technologies.

The philosopher Michael Wheeler cashes out what some of this might mean for our education system. He points out that we already do accept technological augmentation within examinations in maths-heavy topics. We accept that a calculator is OK in some maths examinations (ones involving more complex calculations) but not OK in others (ones involving simpler calculations) because we are testing for a specific, learned ‘on-board’ skill of the students in each case. In the former, more complex calculations case, we accept that to operate at a higher level the subject has to have certain augmentations (the automation of basic workings out): we are thus educating on the presumption that basic calculation technology will be part of their assemblage as a mathematician and so it is appropriate that it be utilised as part of their evaluation. Within the social sciences and humanities I think we should similarly think through what assumptions we should be making about the students we want to create, given the world we live in (more on that later).

Education after AI

Britain has some examples that point to how we can handle this kind of shift already. The Oxbridge tutorial is a small-group supervision that is dialogic by design: this is where a student defends, revises and questions a position in real time. For those universities it’s merely a tradition, but we should revive this across the sector.[12]

During my undergraduate degree in philosophy, I had an oral examination as part of one of my final-year modules; that’s when I first heard that this kind of examination is common across Europe. In our test, we started with basic questions and as we managed to answer them competently, the questions became harder and harder. If we started to struggle, the question difficulty would then reduce slightly in order to keep the conversation flowing and to gauge at what competency level we’re really at.

If you also mix in some evaluation elements regarding group engagement and overall performance across a year, then you start to put in place a culture of learning which is collaborative, which is dynamic, which really gauges understanding and which is not really susceptible to replacement by AI. As Gilman reflects in his article, this will need to be accompanied by an education around asking high quality questions, which students can use to improve their LLM prompts. Learning would then involve working with AI precisely and in an iterative way (not stopping after a single shot), being able to think critically ‘off-AI’ in conversation/tutorial and to interrogate a confident wrong AI answer (which do crop up of course) against the evidence.

Researchers and interlocutors

In light of AI’s likely persistence as part of the technological landscape, and if we follow the logic of cognitive extension developed so far, we are quickly led to the question ‘what are we educating for’? In other words, what kind of educated subject does the higher education process aim to create? What kind of capacities, often augmented by commonly used technological ensembles, do we want to cultivate?

Tentatively, I propose that we should be training students to become two fairly distinct categories of person by the end of their education: a researcher and an interlocutor. It’s helpful to make this distinction – even if it seems somewhat obvious – as it allows us to understand the place of technologies such as AI in a future model.

As researchers, students will be AI-literate, but this will only really have value in terms of its research / literature review capacity. Graduates will leave university being able to find and structure information competently and will be able to understand how to work with AI critically. AI will have little (or only indirect) value when it comes to the student’s grade, given that standardised essay formats are no longer primary (or don’t exist at all) and any research done with AI will be interrogated in conversation with their peers and educators, maintaining the technology as a research assistant and not a replacement.[13] Learning to become a competent, technologically-augmented researcher is ‘entirely appropriate for a world in which the skill of being able to find in real time the right networked information (not just facts, but information about how to solve problems) is arguably more important than being able to attain such information in one’s organic memory’.[14] Wheeler continues: ‘In such a world, which is our world, the brain emerges as a locus of adaptive plasticity, a control system for embodied skills and capacities that enables the real-time recruitment and divestment of technology in problem-solving scenarios’.[15] Being trained as researchers then, is basically training the locus of a hybrid research assemblage which students will take into the world beyond universities.

As interlocutors students will be put in situations largely abstracted from the research assemblage: they will need to be able to retain information, discuss ideas, arguments and develop a thought live with their fellow discussants (be they peers in class or their educator in assessment form). As an interlocutor, we are expressly evaluating the ability to engage without the help of the usual research augmentations, training memory and techniques of argument and discussion. This is entirely consistent with EMT: Wheeler notes that while the fact of cognitive extension means that working ‘externally’ (with tools) is just as much a cognitive process as working merely ‘internally’ with our synapses, ‘nothing in [EMT] rules out the further idea that even though a person’s cognitive system is sometimes extended, we might sometimes want to test the performance of her cognitive capacities under non-extended conditions. In other words, sometimes we might still want to test the naked brain rather than the organic technological hybrid. When this is the case, we will want to ban the use of technology from the examination’.[16] There are indeed such situations where we want to test under non-extended conditions within higher education – and this is because in life beyond university, in wider society including the workplace, it’s still fundamental to be able to engage in argument, discussion, and interrogate ideas as they come at you.

We want a strong set of ‘on-board’ capacities for these situations – we want to be a strong interlocutor – and at the same time we want a powerful capacity to dovetail with our technological extensions (i.e. be a strong researcher). Of course, these two categories – researcher and interlocutor – will often overlap in practice: there could be classes where the lecturer is walking students through an AI-powered literature review; students could produce AI-powered essays, and then the class as a whole discusses the merits of the structure and ways to improve on what’s been produced by an LLM, in a crit-style format; and so on.[17]

Back to ownership

Underneath all of this sits a bigger question of ownership. If AI is now part of the apparatus students think and research with, then whoever controls it holds a lever over how generations to come learn to reason. To leave that to a handful of private platforms, built for engagement and rent rather than understanding, would hand the core of education to firms with their own interests.[18] The tools we think with and through should be treated as cognitive infrastructure: publicly funded, open, and controlled by the people and institutions that use them.

All of the above is an attempt to think through what learning in the social sciences and humanities should be after the fact of AI. Of course, this conversation can’t be abstracted from the economic crisis of universities, their future as an institution and their function within wider society: the fact that the university as such is in question in all of these ways is a good opportunity to rethink the practice as well.

  1. Office for Students, Significant challenges continue to face higher education finances, with nearly half facing deficits in 2025-26, November 2025. Available at: https://www.officeforstudents.org.uk/news-blog-and-events/press-and-media/significant-challenges-continue-to-face-higher-education-finances-with-nearly-half-facing-deficits-in-2025-26/
  2. Times Higher Education, UK university redundancies: latest updates, 2025. Available at: https://www.timeshighereducation.com/news/uk-university-redundancies-latest-updates
  3. Office for Students, Financial sustainability of higher education providers in England: November 2025 update. Available here: https://www.officeforstudents.org.uk/publications/financial-sustainability-of-higher-education-providers-in-england-november-2025-update/ ; Research Professional News, Financial crisis worsens as 45 per cent of English universities face deficit, November 2025.
  4. Higher Education Policy Institute (HEPI) / Kortext, Student Generative AI Survey 2025. Available here: https://www.hepi.ac.uk/reports/student-generative-ai-survey-2025/
  5. Of course, education doesn’t really make sense as a commodity in my view (and maybe that’s the ultimate problem here): the ‘buyers’ of the commodity (who are about 16-17 years old when they make their choices) aren’t really sure what they’re getting, you can’t really take it back if it doesn’t do the trick, and a great many people are actually forced into buying it as a necessary requirement for much of the labor market. Can you say the same things about a flat screen TV?
  6. In this piece I am largely discussing those disciplines which require written essays as one of their main components of assessment (which has been my experience as a philosophy student and then lecturer). This includes a great deal: from the humanities to the social sciences and many other fields. Disciplines such as fine art or medicine (where learning every component of the human body, for example is necessary) or subjects which primarily involve physical technique are less relevant but still often likely require a marked essay system in some form.
  7. Gilman, N. (2026) ‘The University As We Know It Is Finished’. Available here:  https://www.persuasion.community/p/the-multiversity-is-finished?r=1t5vo&utm_medium=ios&triedRedirect=true&hide_intro_popup=true
  8. Will Stronge (2025), What are the policy implications of an updated model of human minds?, The Autonomy Institute; where drawing on Andy Clark and David Chalmers’ Extended Mind Theory.
  9. You can now find Open-source models which can run on a single hefty computer and which are performing as well as Chat GPT-4.
  10. Will Stronge (2026), The ethics of extension: policies of ownership and control for our cognitive apparatuses, The Autonomy Institute,: “the question is not whether to technologically integrate, but in what ways and to what ends.”. https://autonomy.work/portfolio/the-ethics-of-extension-policies-of-ownership-and-control-for-our-cognitive-apparatuses/
  11. Paulo Freire, Pedagogy of the Oppressed (1968; English translation 1970).
  12. Although I have heard that these tutorials often don’t quite live up to their promise in practice, due to how they are structured! Nonetheless the model is there.
  13. This would also save a huge amount of time on marking.
  14. Wheeler (2015), p. 97.
  15. Ibid.
  16. Wheeler (2015), p. 96-7.
  17. There’s an open question that I have not yet resolved, which revolves around the place of writing as a practice and a form of thinking. Writing as thought is, of course, entirely consistent with EMT. The question is: to what extent and in which disciplines do we want to cultivate thought as writing and the skill of thought as writing, and in which disciplines is that less important?
  18. Will Stronge (2026), The ethics of extension: policies of ownership and control for our cognitive apparatuses, The Autonomy Institute,: https://autonomy.work/portfolio/the-ethics-of-extension-policies-of-ownership-and-control-for-our-cognitive-apparatuses/

Will Stronge is Chief Executive of the Autonomy Institute.