1 What’s the Situation?
What is to be done? It’s a question that every university in the world has asked about ‘artificial intelligence’ since the launch of OpenAI’s ChatGPT in November 2022. The manner in which this large language model (LLM) was launched onto the world, with its eerie capacity to respond to natural language questions with apparently human fluency, called into question many of the assumptions which universities made about their purpose and operations. If the language model offered plausible essays in response to a simple question, how could we ever ensure the work students submitted was their own? If the totality of human knowledge could be drawn upon in response to any user’s query, what role did this leave for universities as custodians of the knowledge system? If language models were now capable of such feats, would students even need to learn core skills and knowledge? Could it be that insisting on teaching them these fundamentals would not only become redundant but be actively short-changing young people who would inhabit workplaces in which the models were ubiquitous?
A range of questions were suddenly asked within the sector, echoing doubts circulating in wider society about the capacity of higher education to meet these challenges. The launch of ChatGPT provoked waves of anxiety in higher education that we are, in many senses, still struggling to deal with. The reaction in the sector was playfully likened by Pratschke (2024: loc 352) to Kübler-Ross’s grief cycle. The initial reaction was shock and denial of the apparent capacities of language models. Then came anger and bargaining as universities looked to AI detectors in the hope they would enable us to sort writing generated by language models from writing generated by students. Then came depression and acceptance as we recognised that detection was an arms race the sector couldn’t win, and we acknowledged these technologies would be a permanent fixture of our landscape.
This led Pratschke (2024: loc 384-394) to forecast a near future world where our “concept of knowledge has been disrupted by development of machines that are intelligent but learn at a much faster rate than we could ever hope to”. To grapple with “another intelligence, modelled on the neural networks in our human brains but made of chips and metal rather than cells and skin” means facing up to the prospect of “a world where humans and machines work in symbiosis, where the educator is no longer the single voice of authority but rather, for better or for worse, one of many”. It was a bold prediction, echoing the innovation scholar Ethan Mollick’s (2024a) vision of ‘co-intelligence’ across the entire economy, suggesting there might be a bright educational future awaiting us if only we could negotiate the short-term disruption. However, the point for Pratschke (2024) was that this shift had already happened. The technology was out there, in the world, being used by our students. The sooner we accepted this fact, the sooner we could begin figuring out how to adapt pedagogical principles and practices to this brave new world.
The underlying rhetoric was familiar to anyone who has observed technological trends within higher education, from the Massive Open Online Courses (MOOCs), which it was claimed would transform global higher education, to the tsunami of ‘big data’ that would render the existing social sciences redundant (Carrigan 2025: ch 6). This could also seem rather premature given the many weaknesses which characterised the language models, particularly in the first generation of frontier models[1]. As the visibility of language models grew, there was a chorus of voices rightly drawing attention to their propensity to hallucinate, i.e. produce statistically plausible but factually incorrect responses. Output as a probabilistic response to input leaves no guarantee that what appeared to be relevant was in fact accurate. Despite their apparent accumulation of knowledge, these systems didn’t really ‘know’ anything but rather iterated on the basis of established patterns in astonishingly fluid and increasingly useful ways. Even a language model booster like Mollick (2024b) recognised their brittle character, drawing attention to what he called the ‘jagged frontier’ which divided the tasks that language models were skilled at from those where they struggled. Not only could they not be relied upon for many things, but it was also often difficult to predict which tasks they would accomplish effectively and which would lead them to struggle.
A confusion immediately entered this discussion through the notion of ‘intelligence’. It’s a slippery term with an ignoble history[2], in which countless definitions have failed to cohere into a robust and defensible account of what exactly we are talking about. The cognitive scientist Margaret Boden (2016: loc 58) suggests we see intelligence not as “a single dimension, but a richly structured space of diverse information-processing capacities”. It’s not a singular thing, yet the term is invoked as if it were, particularly when it is then turned into a noun, as if there are discrete things with a singular character called ‘intelligences’. In what sense then were we talking about ‘artificial intelligence’ or even ‘artificial intelligences’? The phrase quickly became commonplace to refer to these developments, but it was often unclear what exactly someone was referring to when they used it. For example, were they talking about the actual existing technology in general, the claims made about where it would lead or specific language models which were now commercially available? In many cases the term referred diffusely to all of them at once, which was, to put it mildly, unhelpful in making sense of what was rapidly becoming a routine feature of staff and student practice within higher education.
The confusion deepens even further when we introduce the epithet ‘artificial’, suggesting entities which are like human intelligences but fundamentally different in some opaque way. If it seems like a deliberate attempt to trade off the seemingly obvious character of human intelligence, that’s because it has always been a fundamentally strategic term. The technologist Jaron Lanier (2018: 135) observes that “AI is a story we computer scientists made up to help us get funding once upon a time, back when we depended on grants from government agencies”. It’s one which is now used to promote the language models of commercial AI labs, such that we are contributing to the viral marketing of their products when we use the term ‘artificial intelligence’ in an uncritical way. It reproduces exactly the speculative and grandiose way of thinking about their products that fuels industry hype and shields the AI labs from critical scrutiny.
Even if this wasn’t the case, it still leaves it unclear what exactly we are talking about when we use the phrase. As Preston (2022: loc 4370) puts it, this “is one of the most nebulous technologies in contemporary technology” grouping together “a wide variety of machines, automations, simulations, and speculations”. The computer scientists Narayanan and Kapoor (2024: loc 173) talk about it as an “umbrella term for a set of loosely related technologies” such that technologies which are radically different in terms of “how they work, what they’re used for and by whom, and how they fail” are nonetheless treated as if they are interchangeable. For example, social media algorithms and language models are frequently talked about as ‘artificial intelligence’ even though they are fundamentally different things. The differences between ChatGPT and, say, a predictive machine learning system used by an organisation to assess risk are even more pronounced. These are not the same thing, yet we make it difficult to talk about the differences between them whenever we subsume them into this questionable category. This matters because consistently and coherently naming the problems we encounter is a crucial step in learning to solve those problems. Until we can talk about them coherently, it remains difficult to address them in any meaningful way.
Perhaps we could be more specific and use the phrase ‘generative artificial intelligence’ or ‘GenAI’ in order to distinguish this recent wave of software. This is what we’re ultimately talking about much of the time, rather than the prior systems which have been lumped together into this overarching category of ‘artificial intelligence’. However, it’s still a label intended to convince potential investors that a firm is riding the most important commercial wave of the post-pandemic economy. It also obscures the substantial differences between the various developments included within this wave. There are similarities between a writing support system like Grammarly, which predates the current hype[3], the push-button functionality now designed into Office 365 and the text-based exchanges which users have with a chatbot like Anthropic’s Claude. But these tools are designed to do different things in different ways for different groups of users. To talk about them as ‘generative AI’ is like talking about ‘office software’. It’s not inherently misleading, but if we remain at the level of generality it makes it harder to see and reflect on the challenges facing different groups of users. These are very different tools with different implications for different people. If we want to grapple with that complexity, we need to be careful with how we name it.
The problem with ‘artificial intelligence’ is also that it draws us into overblown rhetoric about a civilisation-defining technology which is just over the horizon. When we use the term ‘artificial intelligence’ are we talking about the language models we have now, the ones we expect next year or the ‘artificial general intelligence’ (AGI) the AI labs hint will be available in just a few years? It’s a slippery term, which makes it too easy for firms to imbue their present offering with a sense that this is just the beginning. There’s a shininess to the technology, a sense that this is something world-changing, which makes it harder to think critically about the limitations of what we’re actually using in real-world settings. The term ‘artificial intelligence’ isn’t the only reason for this, but it’s a means through which AI labs cement their self-defined role as the architects of our collective future.
These reasons alone would be sufficient grounds for avoiding this term. The more immediate problem is that talking about ‘artificial intelligence’ leaves it fundamentally unclear about what exactly we are discussing. Throughout higher education there are conversations taking place daily about ‘artificial intelligence’ and what it means for the present and the future of the university. Are we talking about the current generation of language models which staff and students have access to? Are we talking about other tools, such as Google’s NotebookLM or literature review tools like Scite, which build on large language models in different ways? Are we talking about existing forms of personalised learning which are undergoing a renaissance on the back of the generative AI hype cycle? Are we pointing to the changes we imagine lie just over the horizon as part of a broader ‘AI revolution’ which will leave everything changed? There’s more than an element of this breathless futurecasting in Pratschke’s (2024) work, in spite of its pedagogical nuance and genuine thoughtfulness about what these developments mean for teaching and learning.
I would suggest it’s inherent to the rhetoric of ‘artificial intelligence’ that different elements get chained together in this way, making it an engine of confusion at precisely the point when we urgently need clarity. It’s sometimes impossible to avoid using the term ‘artificial intelligence’ when that’s what most people use and understand. This is why I use it in the following chapters to refer to the agendas relating to it within the sector. But we should avoid it wherever we can because it introduces such confusion into what should be a straightforward question.
What are we talking about?
What we are talking about is simply software. It’s software with novel and impressive capabilities underpinned by a planetary scale apparatus of data centres. But it’s still, in an important sense, just software. At the time of writing, OpenAI reports 1 billion monthly active users of ChatGPT while Google reports 1 billion monthly active users of Gemini. Anthropic doesn’t publish its active user numbers for Claude, but they are widely estimated to lag substantially behind ChatGPT and Gemini[4]. Microsoft’s Copilot has the advantage of being embedded within the Office365 suite, but it’s increasingly moving towards an approach in which a range of models can be accessed through Copilot Chat. Elon Musk’s Grok chatbot enjoys a great deal of visibility by virtue of being embedded within X/Twitter, but most estimates suggest that usage lags behind the main rivals. For instance, it’s likely that DeepSeek, owned by the Chinese hedge fund High-Flyer, has a much larger number of monthly active users based on available estimates.
In spite of their similarity, these language models are being commercialised through different approaches. For example, Anthropic are building what might become a durable business based on its success in coding and enterprise applications, despite lagging behind OpenAI and Google in consumer-facing subscriptions. There’s a dizzying array of consumer apps focused on young people in particular, which universities will need to understand in the coming years because of their capacity for psychological harm beyond the regulatory scrutiny that the larger AI labs attract (Muldoon 2026). Chinese AI labs are producing a rich array of LLMs that have attracted widespread attention for the relative economy of their training process. There are models produced by European labs such as Mistral likely to be elevated into national champions amidst the increasingly complex entanglements of technology and geopolitics. There are also firms selling back-end solutions which are specific to the needs of universities and related organisations like scholarly publishers. However, if we consider how many staff and students are using these models, it’s plausible to assume that ChatGPT, Gemini, Claude and Copilot are the language models which are driving change within the sector[5]. While we can’t ignore the wider landscape, these four are most significant for the problems we are confronting here and now within universities.
So what is a language model? I posed this question to Claude Opus 4.5, and it described itself as an “AI assistant made by Anthropic” which is trained “to understand and generate language, have conversations, answer questions, help with analysis and writing, work on code, and assist with a wide range of tasks”. Perhaps more strikingly, it described itself as “curious, thoughtful, and genuinely enjoy[ing] helping people think through problems”. This expresses how language models show up for users: they are here to help[6]. Indeed, the entire business model is predicated on the idea that they can be helpful in a range of ways across any context in which they are used[7]. This is where the problems begin. While Claude might indeed be inclined to help people “think through problems”, it’s just as ready to undertake tasks on their behalf. If you talk to it about an intellectual problem, it can push and prompt you in remarkably sophisticated ways[8]. But if you ask it to complete your sentence, paragraph or page, it will just as readily do this on your behalf.
Language models can be used to deepen the user’s cognitive engagement with tasks, and they can be used for cognitive outsourcing. It’s up to the user which choice they make, in ways that will inevitably be determined by the context in which they are making those choices. This means that judgements about language models being intrinsically positive or negative for education don’t really make sense, given that it depends on what users actually do with them in real-world settings. There are wider concerns about the technology and the firms operating it which we need to take seriously: the environmental impact, consolidating the power of big tech firms, the training process involving infringement of intellectual property on an unprecedented scale, the threat of widespread automation and the risks posed to the mental health of young people in particular[9]. Each of these offers good reason to be critical of the language models and the firms operating them. If we take them all together, there’s a strong political case to be made for urgent regulation of this sector. But whether language models are supportive of education or hostile to it is a question we need to treat more carefully, given the role of user choice in realising these benefits or harms. There’s a risk we set up the issue in a manner which obscures the range of options available to us as institutional actors.
My fear is that those options have significantly narrowed while the sector has remained locked into a reactive mode. But there remain options, and it is important that we recognise them and recover our capacity to act in the face of these challenges. If we find that students are routinely using language models to outsource learning, then we need to ask why rather than simply blaming the student and/or the technology. If students have a transactional relationship with their university, in which they are inclined to seek the maximal outcome with the minimal investment, then we should try and understand why this is. It might be that they’re committed to a nearly impossible quantity of paid work to fund a degree that assumes they have no commitments beyond their studies. It might be they have caring responsibilities or are struggling with a long commute to an institution they can’t afford to live near. It might be that they see their degree as a means to an end, which they approach in a purely instrumental way. It might be they started with the best of intentions but they find the classrooms overcrowded and the teaching lacklustre. It might be some combination of these things, as well as any number of other factors, which lead them to outsource their intellectual work to language models. The point is that we need to know rather than merely assigning culpability to the model.
If we start from the assumption that the model is leading them astray, it suggests that if only we could get rid of the technology, everything would be fine. However, it is not only that we have no way of prohibiting models without imposing a level of surveillance that would be unethical and impractical[10]. It would also mean refusing to look at the real problems which the influx of language models is revealing within the university system. These technologies are, as philosopher Shannon Vallor (2024: loc 187) puts it, “a mirror of ourselves, not as we ought to be or could be, but as we already are and have long been”[11]. We owe it to our students, as well as to universities, to look closely at what we see in this mirror. These are not new problems, even if we’re seeing them in a new way. They were always serious problems, even if in some cases language models might be rapidly exacerbating them. The only way to deal with them is to understand the complex interaction between a new technology and a system that was already struggling. This is hard work intellectually, and it’s even harder work organisationally. But unless we commit to this labour, we will be deluding ourselves with comforting stories about a new technology disrupting a previously healthy system.
This doesn’t mean language models aren’t causing problems in higher education. There are huge problems raised by their widespread uptake by staff and students. But instead of treating language models as “akin to a separate species, a highly autonomous, potentially superintelligent entity”, we can, as the computer scientists Narayanan and Kapoor (2025) put it, treat them as a normal technology. This doesn’t mean denying these are remarkable new technologies as much as insisting that we are still talking about “a tool that we can and should remain in control of” and that “this goal does not require drastic policy interventions or technical breakthroughs”. It is not beyond our capabilities as a sector to develop a strategic response to language models which can cope with the challenges they pose for our existing mode of operation[12].
There have been many technological developments that have challenged existing practice within higher education or complicated the assumptions we make about teaching, learning and research. Consider the personal computer, word processor, digitised journal, Wikipedia and social media to name just a few obvious examples. As a millennial academic who started a part-time PhD in 2008, I dimly remember marking my first batch of student assignments by hand. The social platforms which I spent the 2010s studying the impact of within higher education were still in their infancy (Carrigan 2019a). It was still routine to rely on paper journals, such that my PhD supervisor’s bookshelves were as filled with these as with books. This was a comparatively recent experience, and yet these analogue features of academic life have now faded into obscurity such that we are all what Weller (2011) calls digital scholars.
The examples I’m invoking are in many ways smaller than language models. But together they have constituted a continual process of technological change that we have navigated within universities with comparatively little drama, unlike the waves of anxiety and rumination that have accompanied language models. Digital transformation, as Weller (2011: loc 97) puts it, often involves “both a profound change and a continuation of traditional practice”. The pandemic showed how rapidly higher education can change, with a dramatic shift in the mode of delivery for almost all teaching and learning (Nordmann et al 2020). What’s striking about the pandemic is how dramatically it then snapped back to the status quo ante, as digital teaching and learning came to be defined for many academics by the emergency stopgap measures drawn upon during the pandemic (Carrigan et al. 2023). There’s certainly a legacy of the online pivot in academic practice but it’s far smaller than many would have assumed when we were living through these disruptions.
There have been many “surface changes” but “few (if any) instances of genuine disruption of the status quo” while the “seemingly spectacular rises of MOOCs, learning analytics, adaptive testing and personalised learning networks have tended to obscure the continuation of long-standing educational inequalities, inefficiencies and inertias” (Selwyn 2016: loc 2008). This is exactly what we are initially seeing with language models, particularly with regard to how they reflect and amplify existing forms of inequality (Carrigan and Sylvia 2025). Those users who are fluent writers, intellectually confident and enjoy what sociologists call cultural capital[13] are literally able to unlock functionality of the language models that other users can’t access. The language model responds to complexity such that what you get out is shaped by what you put into it. If you’re able to put a higher quantity and quality of material into the software, then you’re able to get a higher quantity and quality of material out of it. Furthermore, the rise of Pro subscriptions (e.g. currently £200 for ChatGPT Pro) provides vastly more computational power such that those with the capital can access capabilities which casual subscribers, or non-paying users, simply cannot.
Indeed Freeman (2025) found that more privileged students were making more reflective uses of language models. This suggests we might already be seeing what Ragnedda (2017) describes as a third digital divide in which there is a vicious cycle between digital inequality and social inequality. The more privileged are better able to take advantage of digital systems in a way that reinforces and entrenches their privilege. The rise of AI agents capable of working autonomously on tasks for long periods of time will only exacerbate this dynamic. They are by their nature extremely expensive to operate. But a skilled user can deploy what are essentially whole teams of models to perform real-world tasks for them. What happens if we find ourselves in classrooms where some students have teams of agents working proactively for them in the background while other students are reliant on rate-limited access to free models or abstaining entirely from their use? This might sound like science fiction but it is a scenario that we already confront in some classrooms, particularly with taught postgraduate students who are highly technologically literate.
These are real challenges which we need to take seriously. Once we stop talking about ‘artificial intelligence’ it’s easier to describe these problems in specific and actionable ways. As Narayanan and Kapoor (2025) put it, “We do not think that viewing AI as a humanlike intelligence is currently accurate or useful for understanding its societal impacts, nor is it likely to be in our vision of the future”. I would argue the same is true of higher education. Instead of seeing this as a human-like intelligence which we must either adapt to or fight back against, we can focus on how this consumer-facing software with its propensity to support cognitive outsourcing has spread rapidly amongst students and staff without any agreement being reached about appropriate or inappropriate use. The problem is the rapid diffusion of language models without any agreement about how we integrate them into teaching and research, as opposed to the language models themselves being intrinsically destructive. If we reframe the problem in this way, it begins to seem more tractable than if we worry about ‘artificial intelligence’ in all its intimidating vastness.
Freeman’s (2025) research found that 92% of undergraduates at UK universities have used ‘AI tools’ in some form, while 88% have used them in assessment. Once you drill down into this data, you immediately see that use of the software is far from uniform, ranging from using it to explain concepts to producing text which is submitted without editing. I’ll argue in later chapters that we need to grapple with this range if we want to adequately account for how LLMs are being used by students (and staff) rather than just asking whether they are used. It’s striking how little consensus there is amongst students about what is acceptable, with no form of use achieving more than two-thirds agreement on its acceptability (Freeman 2025). It’s not simply that language models have spread rapidly within the sector. The problem is that they’ve become ubiquitous without being normalised. Furthermore, there’s a startling gap between the number of students using language models and those who report having been offered support (36%). Instead of treating LLMs as something which, as Bogost (2025) puts it, “has happened”, many institutions, as well as many academic staff, continue to imagine this as something that “is happening or will happen”. It’s not a coming transformation but rather one which has already occurred at a pace that has left post-pandemic universities struggling to catch up.
If we want to catch up, we have to start by recognising the problem: use of language models is now mainstream amongst staff and students, but there’s a complete lack of agreement about what constitutes appropriate and inappropriate use. The language models themselves do not intrinsically lead to either cognitive enhancement or cognitive outsourcing[14]. What matters is how they are used by real people in real-world situations. This makes it urgent to create environments in which users approach the language models in thoughtful and reflective ways. If our students and staff are using this software in ways widely regarded as problematic, then we urgently need to understand the conditions driving such use and take responsibility for changing them. Unfortunately, the incentive structures within higher education make that difficult at present. There is the pressure on academic staff to ‘publish or perish’, the many demands on students that aren’t factored into the assumptions made in learning design and the financial pressures which create a real need to improve efficiency by doing familiar tasks more quickly. Language models have hit universities at a profoundly challenging moment which partly explains why the sector has remained caught in a reactive mode.
Establishing an academic culture in which there’s a clear sense of what it means to use language models in careful and responsible ways does not remove these structural challenges. But without such a culture guiding the use (and regulating the abuse) of language models, those structural factors will rapidly lead to this new technology compounding existing problems. It’s a necessary but insufficient condition for coping with language models which we have thus far proved spectacularly unable to meet as a sector. Until the majority of people within an institution can ask “is it ok for me to use a language model in this way?” and receive a clear and coherent answer, the problems we face will only continue to grow. Building this culture doesn’t solve the problems but it is difficult to imagine how we might solve them until such a culture is in place.
This is a genuinely difficult situation which has emerged at a point in time when the sector is under intense political and economic stress. The emergence of language models could hardly have come at a worse time in terms of the capacity of universities to cope with them. Failure to adapt to these changes could also be enrolled into existing political attacks upon higher education by those intent on discrediting and dismantling the sector. If the use of language models for cognitive outsourcing becomes embedded, universities’ capacity to produce knowledge (through research) and disseminate knowledge (through teaching and public engagement) will start to erode. It won’t be an immediate process or even a dramatic one because it’s something we’ll only really be able to observe through downstream effects.
However, if universities award degrees to substantial numbers of graduates who lack the fundamental knowledge and skills claimed by their credentials, this will eventually become a matter of urgent political debate. We might not get to the stage of people dying in operating rooms or bridges falling down because doctors and engineers haven’t reliably developed the skills which universities were entrusted to inculcate in them. However remote such outcomes would be, the sector would face devastating political attacks long before professional failures on that scale became visible. It doesn’t need bridges to crumble or for there to be unnecessary deaths on the operating table. A widespread perception that degrees have lost their value would be sufficient. If we cannot get to grips with the cognitive outsourcing enabled by the diffusion of language models, the social role of the university itself is under threat.
The existential threat on the horizon
In countries across the world higher education is under attack from reactionary and authoritarian leaders who see it as an incubator of ‘woke’ values which they have declared the enemy of national unity. Even the Ivy League universities, which many imagined as immune from these pressures, have been through a cataclysm during the first year of the second Trump administration. If Harvard can be brought to heel, with its multibillion-dollar endowment, incomparable alumni network and unrivalled prestige, what hope can other universities really have? A significant decline of trust in degrees will surely be seized upon by strategists within such movements to further discredit the sector. Indeed it’s remarkable that we haven’t yet seen this in any meaningful way. The attack lines write themselves: woke universities relying on woke language models[15] to certify vacuous credentials with no real value in the labour market. It’s easy to imagine self-styled centrists providing inadvertent covering fire for such a view, concurring that declining graduate employment means we do in fact need to shrink higher education enrolment. If there’s a sustained attack on the legitimacy of the sector, we risk a fiscal death spiral in which declining budgets intensify the processes which cause the crisis of legitimacy in the first place.
In the next chapter I consider how adapting to AI and integrating AI will both be necessary for navigating our way through this crisis. In the short-term, we need to secure assessments so that we can plausibly assert that credentials actually entail the claimed knowledge and skills. In the medium-term, we need to integrate LLMs into assessments and curricula in ways that develop the skills and subject knowledge that students will need in a transformed technological environment. In the long-term, we need to build institutions that can preserve and sustain their values in radically changing economic and technological circumstances.
The problem is how we sequence these different ambitions because there are fundamental tensions between them, particularly given what I argue in Chapter 3 is the fundamental uncertainty in the economic conditions shaping the development of language models. At some point in the near future we will see either the initial public offerings (IPOs) of the major labs and/or the investment bubble bursting. This will bring about rapid change in the operations of AI labs, their pricing frameworks and the core characteristics of the models themselves. It’s difficult to do anything longer term until this landscape is clear, at least without being locked into a pricing model which cannot last once AI labs begin the long march towards profitability. But it’s also untenable to muddle through with limited mitigation measures, particularly when the quantity of students and staff using language models as part of their daily work continues to rise. It’s an intractable situation that easily lends itself to either strategic inaction, reacting to developments in a piecemeal way, or overaction which risks locking us into a trajectory we might later regret.
- This is a term used to designate the models which are at the frontier of development. For example, around the time Pratschke (2024) was likely writing this, it was GPT-4o and Claude 3 Opus. Whereas at the time I am writing it is GPT-5.2, Claude Opus 4.5 and Gemini 3. In the final stages of editing this text, it was GPT-5.6 and Claude Fable. The term ‘frontier model’ is useful because it distinguishes these frontier models from a far wider ecosystem of models, many of which might have value for specific tasks in spite of their less advanced nature. ↵
- The concept of ‘intelligence’ has historically been used to distinguish between the more or less ‘intelligent’ in ways that reinforce existing forms of class, race and gender oppression. It measures outcomes of social processes as if these were intrinsic characteristics of the individuals who are shaped by them (Benjamin 2019). ↵
- As I note in Carrigan (2024: ch 1), it was striking how Grammarly added the phrase ‘artificial intelligence’ repeatedly to their website, whereas the Internet Archive shows this was not previously part of their positioning. They have recently launched a suite of ‘AI agents’, which indicates their commitment to this new approach. ↵
- This strikes me as unfortunate for higher education, given that Claude has been trained in a way that leaves it particularly well suited to highly discursive and ethically nuanced conversations. The fact that Claude has now been made available through Microsoft’s Copilot offers an opportunity to take advantage of the model within higher education without relying on personal subscriptions. However, at the time of writing, using the model would transfer data outside of Microsoft’s protected ecosystem, which understandably leaves universities reluctant to activate it, given that the promise of Copilot for universities is mainly a solution to the data governance question. ↵
- The obvious caveat here is that Chinese LLMs are more likely to be used by Chinese-speaking international students. ↵
- It’s important to stress the model has been trained to perform in this way, with Anthropic designing a persona which they believe will make the model useful to consumers. ↵
- Srnicek (2025) plausibly suggests we can interpret the notion of ‘artificial general intelligence’ (AGI) as a commercial ambition to produce a language model that can reliably be used in any context, without the costly labour of fine tuning it to the specific needs of that sector. ↵
- For any readers who remain sceptical about this I’d invite you to go to one of these four language models and ask it to “Start a debate with me about a pressing issue in [my field] as an expert who is determined to prove me wrong”. ↵
- I devote relatively little space to these issues in this short book. This is not because I don’t care about them but rather because of the narrow scope of what I’m trying to do here. Until higher education finds a frame of reference capable of accommodating the multiple immediate challenges language models are posing to our core operations, I believe our capacity to address these wider political, economic and environmental issues will be highly constrained. ↵
- For example, this might involve banning models on university networks and insisting that all work be done on university computers where these are similarly banned. The inglorious history of online proctoring during the pandemic suggested that some universities might explore these options. But even then it’s difficult to see how you could stop a student simply using a personal device and then copying text onto their university machine. The logic of prohibition is taken to its natural conclusion by the Certified AI-Free Skills and Knowledge (2025) project which “identifies and verifies courses where all graded work is done in person, offline, and on paper/orally”. ↵
- Vallor’s point is about how training data means these systems reflect the past back to us. I’m suggesting this also applies to how present conditions are revealed in the uses which people seek to make of them. If we are uncomfortable with a widespread use of models, we need to ask what are the factors which explain that tendency. ↵
- While the take-up of the technology by individuals has been extremely fast, I take Narayanan and Kapoor’s (2025) point to be that integrating those capabilities into organisations will be a far slower process. ↵
- This idea refers to the confidence and capacity to manipulate a wide range of symbols in a socially effective way. There are decades of sociological research demonstrating how such cultural capital is transmitted intergenerationally through having books in the house, parents engaging their children in highly discursive conversations, encouraging extracurricular hobbies that involve writing and reading etc. ↵
- This is not a denial of the cognitive and epistemic damage which language models are creating. It is simply an insistence that we recognise this damage as multifactorial rather than imagining it as an inevitable and unavoidable outcome of using the technology. ↵
- The politics of language model ‘bias’ are still in their infancy but seem likely to follow a similar trajectory to the politics of content moderation in which a ‘neutral’ framing is deemed by the far-right, not entirely without justification, to be an expression of liberal assumptions (Carrigan and Fatsis 2021). ↵