Introduction
Large language models (LLMs) first appeared to the general public as a curiosity. Today, everyone uses them, and there is no doubt that their use is radically transforming the nature of our technical activities. This is particularly true in mathematics. Progress seems exponential (the unit distance problem, the Jacobian conjecture, , the Navier–Stokes Millennium Prize Problem), and it is difficult to see why it would hit a ceiling. I have personally followed developments in AI throughout my PhD in hydrodynamic instability, mainly focusing on the use of agents for coding.
I have not worked in pure mathematics for a few years now, but the reflections that follow extend across mathematics as a whole. I would also like to note that the stakes seem different to me in pure and applied mathematics; I will return to this.
In this article, I offer my interpretation of a few arguments put forward by two philosophers of technology, Günther Anders and Bernard Stiegler, in the context of the development of AI in mathematics. Drawing on my own experience and on these writers, I propose a framework for understanding what is happening to us, and introduce what I consider to be the right concepts for thinking and acting in this new environment.
The Promethean Experience: Anders, the Gap, and Shame
AI is moving too fast. That is a cliché. At the macro level, of course, we can see that AI is advancing extremely quickly, faster than we know how to regulate it (see the recent OpenAI–Hugging Face incident). At the level of an individual user, it simply thinks too fast for us. It races ahead. It types too fast, codes too fast, thinks and draws increasingly accurate conclusions too fast. In other words, our capacity to absorb and understand cannot keep up. [5]
Günther Anders was a post-war philosopher of technology. The revolution of his time was the atomic bomb. For him, Hiroshima marked the moment when humans produced, for the first time, an object of which they could no longer form a concrete mental picture. This was as true morally as it was physically. The damage inflicted by the bomb was so immense that it was virtually impossible for a human mind to comprehend it in all its dimensions: in joules, in lives, in consequences. There is too much information in too little time: a gap between what humans produce and their representation of what they produce. Anders calls this the Promethean gap, in reference to the myth of Prometheus.
In The Obsolescence of the Human, he tells the story of a pilot who, after bombing towns during the war, tries to take stock of what he has done. The soldier retreats to a monastery but, a year later, still cannot fully grasp the significance of his actions. He has carried out acts that his conscience cannot catch up with. There has been a lag between the speed at which technology enabled the crimes to be carried out and the human capacity to come to terms with them. [1]
I think we are experiencing a similar lag with artificial intelligence, in every field, and particularly in mathematics. It is a curious coincidence that Hugo Duminil-Copin, in his post here, used precisely the word “nuke” to describe AI’s effect on the mathematical landscape, given that Hiroshima is Anders’ focal point. [4]
The technical object leaves us in the dust. It is too fast. Or rather, we are too slow. Writing in 1956, Anders observed that a defining feature of the industrial age was the shame humans felt at not being machines, as though Icarus’s wings might have reached the sun had they shed their human burden. Humans cannot match machines’ performance and must therefore adapt as best they can so as not to hold them back too much. To return to the military example, fighter pilots have to train and adapt their bodies to withstand the high g-forces generated during manoeuvres. Likewise, mathematicians seem to need to develop their ability to run tasks in parallel, skim their agents’ verbose walls of text, and make research decisions quickly. To be more specific, they need to adapt if they adopt a productivity-driven approach and wish to compete with AI-boosted mathematicians in publication output. Which is not obvious.
Without indulging in armchair psychoanalysis, I think shame is indeed present, unconsciously, in many of us. I think we all feel a certain admiration for the power of these LLMs and therefore, in a sense, regret that we are not as quick-witted. Personally, I struggle to keep up with my Claude’s reasoning. And I feel a bit awkward answering its questions so that it can make progress on my codebase, as though our roles had been reversed.
Stiegler and Proletarianisation
I would also like to discuss Bernard Stiegler. He is not particularly well known in France, and is probably no better known abroad. His extraordinary life took him from robbing banks to discovering philosophy in prison, studying under Jacques Derrida, defending a doctoral thesis at forty, becoming a leading figure in his field, and putting his concepts into practice through collaborations with industrial groups.
In Automatic Society, Stiegler revisits the concept of proletarianisation. For Marx, the proletarian is the worker who loses his know-how to large-scale industry. Whereas the weaver of old embodied his know-how in his gestures, he finds himself “reduced” when the mechanised loom appears. The machine performs the gesture; the worker supervises. His knowledge has been expropriated and transferred into the technical object. He has been forced to unlearn. In the interests of economic profitability, he has been enslaved to a machine and alienated by being deprived of his identification with the objects he produces. [2]
Stiegler generalises this idea: proletarianisation is the loss of knowledge in the broadest sense. He describes three waves of this loss. First, knowing how to do (Marx, industrialisation); then, knowing how to live (consumer society, in which marketing takes charge of our ways of living); and finally, he envisages the loss of knowing how to think (our cognitive and theoretical faculties) to ever more intelligent digital objects.
I do not think I am mistaken in saying that LLMs and AI agents, particularly when used intensively, rapidly, and in parallel, embody this movement. LLMs have removed our need to learn to write properly; they can act as extremely accurate and fast translators, and they understand nuance. They serve as psychologists, tax and legal advisers, substitute parents, and teachers who are always available. They debug code far more efficiently than humans. They find mathematical proofs more quickly; they calculate faster. By centralising all these services in a single tool, they become an instant cure-all answer to more or less any question, even the hardest ones. In short, they serve as an oracle for the most complex tasks as well as the simplest.
In this sense, it is tempting to incorporate a layer of GPT into our own thinking process. We might replace our judgement with AI’s when making difficult decisions, where the stakes are high enough to justify the intervention of a superintelligence. Or, conversely, we might automate quick decisions and small tasks because they are not worth our human attention. We might do one or the other, or both.
I am not saying this is good or bad. I am saying that it is going to happen, and that it is happening already. AI supplanting our capacity to reason is not merely a looming threat. It is already under way.
Why We Do Mathematics
At this point, I think it is important to take a detour through the question of what gives mathematical activity its meaning. If we consider what we lose and what we gain from AI in mathematics, we must assess these gains and losses in terms of how they serve our motivations for doing maths.
At the most recent meeting of Cambridge’s mathematics department on the use of AI, one comment that came up was the following:
I do not do maths to produce papers. I do maths because I enjoy it. So I do not want to deprive myself of that enjoyment by using AI.
From conversations with those around me, I have come to think that every mathematician is driven by two motivations, in proportions that vary from one person to another:
- The pleasure of mathematical activity in its broadest sense: getting stuck, searching, discovering, marvelling.
- The pleasure of advancing knowledge, obtaining results, and perhaps even having an impact in the real world.
On average, I think the first motivation is probably more prominent in pure mathematics, and the second in applied mathematics. And it seems to me that these two branches are also broadly guided by the former and the latter, respectively.
Avoiding Unlearning
But in both cases, and this is just as important, we all practise mathematics as a human activity that defines us and individuates us. Individuation, a concept Stiegler takes from Simondon, refers to the process through which an individual comes into being: an always-unfinished process of becoming, in which one is transformed through contact with their surroundings and with others. The product of mathematics is not only the set of proven results; it is also the mathematician. We do not merely do mathematics: by doing it, we make ourselves mathematicians.
Drawing on Plato’s Phaedrus, Stiegler takes up the idea that technology is both remedy and poison, and that this tension is constitutive of humanity’s relationship to the world. Humans need technology to become themselves. As cultural individuals, they develop their singularity by appropriating a collective inheritance of knowledge and practices, made accessible through technical tools and media that preserve memory. But the same technology can also make them “less themselves”; that is, it can disindividuate them when it replaces the practices and efforts through which they make knowledge their own and learn. The same technology can enhance them or diminish them. Technology is neither inherently good nor bad. However, a good use of technology is one that connects them to others and strengthens their individuation. A good use of technology is one that extends human capacities rather than diminishing them. It is one that creates “circuits” of individuation: structures (research groups, universities), collaborations, careers, and progression, rather than short-circuiting these processes by turning individuals into consumers instead of active participants. [3]
This is not an entirely objective criterion, but I think we can readily tell when a technology dulls our minds, enslaves us, or diminishes our abilities, and when it develops our skills, our character, and our relationships, allowing us to learn.
As the field undergoes radical transformation, the challenge, in maths as elsewhere, is not to unlearn. For many of us, GPS has supplanted our ability to find our way rather than improving it. I think that, if we still want to do mathematics, rather than simply avoiding AI, it is important to ask ourselves whether using it is causing us to unlearn the subject.
As a graduate student, one of the best uses I have found for AI is having it write lecture notes. Whenever I want to learn something, acquire the theoretical background needed to understand a paper, or develop skills, Claude Code writes clear, instructive lecture notes for me, which I then study. I find this enjoyable and effective, and it seems to remove the wrong kinds of friction from the process of understanding while preserving the right ones.
My reason for mentioning this example, beyond sharing it, is to show that we do not have to choose between AI and learning. The same tool can enable either learning or unlearning, depending on how we incorporate it into our cognitive processes.
Conclusions
It seems to me that doing maths is only worthwhile if we are not “proletarianised” in the process; that is, if we can:
- Identify with the activity, whether as professional work or as a hobby.
- Interact with a community, locally or globally.
- Continue to learn.
- Enjoy it.
In conclusion, I would like to raise a few questions to which I do not really have answers, but which suggest directions for debate:
- How can we use AI to learn rather than unlearn?
- What sustainable funding model could protect mathematicians and their activity, in both pure and applied mathematics?
- From an industrial perspective, how can we adapt to the inflationary devaluation of mathematical work (see, for example, Claude formalising Fermat’s Last Theorem in Lean far more quickly than the human team working on it)? [6, 7]
- What is the nature of the difference between having AI prove a theorem, proving it with limited AI assistance, and proving it entirely on one’s own? Can we be content with recalling proofs and rediscovering them as exercises, or is there a particular pleasure in finding them on our own for the first time? Can we do without that pleasure?
I think it is worth reading (or rereading) Stiegler and Anders, who, long before the development of AI, reflected on questions that are so relevant today. I hope I have at least sparked your curiosity about them. I think we need interpretative frameworks and concepts to think through what is happening to us.
References
1. Anders, Günther. The Obsolescence of the Human. Originally published in German in 1956.
2. Stiegler, Bernard. Automatic Society, Volume 1: The Future of Work. Originally published in French in 2015.
3. Plato. Phaedrus.
4. Duminil-Copin, Hugo. “Care for a Little More AI?” Proofs and Prompts, 30 August 2026.
5. OpenAI. “The Hugging Face Incident and the Road Ahead.” 26 August 2026.
6. Anthropic. “Formalizing Fermat’s Last Theorem.” 4 September 2026.
7. Buzzard, Kevin. “The Fermat’s Last Theorem Project.” Lean Community Blog, 2024.
Crossposted from my blog.
Received 14 September 2026.
Add to the discussion