Written for Proof and Prompts. For context: Proof and Prompts is a communal blog for mathematicians’ thoughts on how AI is changing mathematics.
This essay is not a high-level reflection on what AI means for mathematics. It is more of a “what should I do with my life?” essay. As such, it may not be well suited to ProofAndPrompt, but I needed to write down my thoughts (for what they’re worth) and to mourn, somehow.
I had a depressing August. Like where I compulsively bought stupid stuff every time a conjecture was solved through AI by a Redditor who didn’t understand the solution, probably not even the question. I, then, started doing it myself. There was a statistical problem that I found interesting and had had in mind since my PhD, but had never had the time to tackle. In a few hours, I obtained something reasonably decent. With a bit of polishing, it could have made it into a good journal. Before, it would have taken me years… But why bother putting this on arXiv? Why send it to a journal? I can’t accept someone spending more time reading this than I spent producing it. But the main point was this: work that should have taken me a long time and a lot of effort could be done almost instantly. I told a colleague, “I didn’t think my job would be the first to be replaced.” I was joking, but there was still something real in it.
The loss began far earlier for me. I was never a top-level mathematician, but I always hoped to produce some impactful and interesting results, and worked hard to achieve it (a hope that is gone now). At least, among my colleagues at the lab, I believe I had some value. Because there were topics I knew very well, because I was reasonably technically competent, and also because I was very good at coding. I could produce complex simulations that helped our projects find research directions or ideas. This ended more than a year ago. I think I haven’t written a line of code for a year and a half. By then, everyone in my team could have produced their own simulations. Interestingly, somehow, nobody does (at least as far as I am aware). That was already a loss, but at the time I naively thought I was still a decently competent colleague who could help my teammates with the maths. Now that is gone too.
Then September showed its face, the time when most colleagues come back from holiday. I couldn’t wait to share my depression. I was a bit disappointed that almost none of them saw the wall coming as I thought I did. Then there were the $\theta(p_c)=0$ rumours and the leaked Anthropic GitHub commit, and it began to feel a bit more real to everyone. But still, I didn’t feel that my colleagues (apart from the PhD students I spoke to) realised what they could do themselves on their computers.
Some colleagues say that this will only be transitory, that the gains from using LLMs will quickly be outweighed by the costs. I think the opposite. The first time I was impressed by an LLM was with DeepSeek in January 2025. I gave it a pretty hard undergraduate analysis problem, and it managed to solve it. This wasn’t worrying at the time. My point is that DeepSeek needed data center grade hardware. Now models that can solve this kind of problem, even faster than DeepSeek, can run on my computer (a high-end consumer computer, but still). My guess is that models will become far better than they are now, and at a fraction of the cost (in any units you like).
I have a few ongoing projects, and I am pretty sure they could nearly all be solved in one shot in half an hour with the best version of Astra. I could do that, polish the result, send it off, and produce an article in a month that would not have to fear the comparison to my earlier productions. I believe people are already doing this[1]: I was looking at math.PR on Monday and thought, “That’s a lot of single-author articles today!” So I tasked my favourite agent with investigating how the number of single-author articles had changed, and I was not surprised by the result[2].
The share of single-author articles seems also to increase[3].
(Be re-assured: no humans were harmed in the production of these plots. In particular, none had to work or even think—most importantly, not me…)
I could finish every single project I have in a few weeks. I won’t do it because I still have some affection for these problems. But once they’re settled, I don’t think I will ever write an article the way I did before.
So what should we do?
First and foremost, I really think the most important step we should take as a community is to get rid of the idea of “ownership” of mathematical results altogether. It doesn’t make any sense to me now. Is it your theorem if AI proved it? Is it your theorem just because you did it first, if I could obtain the same result in ten minutes with Astra? To be honest, the problem existed well before: how many people have been listed as authors just for asking the question? That doesn’t feel so different to me from prompting an LLM.
It is a hard step, because it is so ingrained in us. Our whole professional life is built on the results we have fathered. Our whole idea of ourselves as mathematicians is constructed around it. All our myths, all our culture. Maybe I am not well placed to say this, because I’ve always been a mediocre mathematician, but look at the dispute over Navier–Stokes, fighting over who prompted the LLM better or first [4]. Who will be king of the ashes? I think it doesn’t matter anymore and, most importantly, it shouldn’t matter to us.
One could argue that disclosing AI use can solve the problem of ownership, but it is not verifiable. And I don’t trust people in a competitive environment. Anyway, the culture of priority in mathematics has always been toxic. If we get rid of it, I believe we can all be happier.
Concerning the practice of mathematics, I’d say, “you do you”. We don’t own mathematics, we don’t own the problems. People shouldn’t be ashamed of using AI.
On the other hand, if you find meaning in doing everything by yourself, good. It is just not the way I am built. I see no sense in spending months working on something that could be so easily solved by chatGPT. I can’t accept putting my name on an article that relies substantially on AI for doing the maths, either. That makes no sense to me. Maybe I will try to do mathematics in the following way. A colleague recently talked to me about the question of the large particle number limit of a two-dimensional log gas confined to a disk. The result is known only in a very particular case where the process has a determinantal structure. I think I will create a GitHub repository containing the results known either from the existing literature or through AI. Anyone will be welcome to contribute in any way they want. The goal would be to produce a coherent, digested account of the mathematics which, once it covers the question sufficiently, could be put on arXiv in the name of a team.
I would like to end with a word for PhD students and postdocs. If you, reader, are lucky enough to hold a permanent position, please be aware that they are probably far more aware of this situation than you are. In particular, if you disagree with me and think we are not going through a terrible transition, please acknowledge at least that they have good reason to feel scared and lost. Please allow them to use AI in any way they want, or not at all if they don’t want to. And more importantly, if you are involved in recruitment of any kind: stop counting candidates’ articles. Whatever people say, this has always been a very important criterion. It was already a bad one before, now it is just terrible. It creates a very wrong incentive. They need to hear that they don’t have to use AI to remain “competitive” and the community must acknowledge it.
Between August 2025 and August 2026 the number of single authored articles went from 1,174 to 2,603. Among the 2,603 articles of August 2026, only 809 mentioned AI (take these numbers with caution, they are uncertain)
Written for Proof and Prompts. For context: Proof and Prompts is a communal blog for mathematicians’ thoughts on how AI is changing mathematics.
This essay is not a high-level reflection on what AI means for mathematics. It is more of a “what should I do with my life?” essay. As such, it may not be well suited to ProofAndPrompt, but I needed to write down my thoughts (for what they’re worth) and to mourn, somehow.
I had a depressing August. Like where I compulsively bought stupid stuff every time a conjecture was solved through AI by a Redditor who didn’t understand the solution, probably not even the question. I, then, started doing it myself. There was a statistical problem that I found interesting and had had in mind since my PhD, but had never had the time to tackle. In a few hours, I obtained something reasonably decent. With a bit of polishing, it could have made it into a good journal. Before, it would have taken me years… But why bother putting this on arXiv? Why send it to a journal? I can’t accept someone spending more time reading this than I spent producing it. But the main point was this: work that should have taken me a long time and a lot of effort could be done almost instantly. I told a colleague, “I didn’t think my job would be the first to be replaced.” I was joking, but there was still something real in it.
The loss began far earlier for me. I was never a top-level mathematician, but I always hoped to produce some impactful and interesting results, and worked hard to achieve it (a hope that is gone now). At least, among my colleagues at the lab, I believe I had some value. Because there were topics I knew very well, because I was reasonably technically competent, and also because I was very good at coding. I could produce complex simulations that helped our projects find research directions or ideas. This ended more than a year ago. I think I haven’t written a line of code for a year and a half. By then, everyone in my team could have produced their own simulations. Interestingly, somehow, nobody does (at least as far as I am aware). That was already a loss, but at the time I naively thought I was still a decently competent colleague who could help my teammates with the maths. Now that is gone too.
Then September showed its face, the time when most colleagues come back from holiday. I couldn’t wait to share my depression. I was a bit disappointed that almost none of them saw the wall coming as I thought I did. Then there were the $\theta(p_c)=0$ rumours and the leaked Anthropic GitHub commit, and it began to feel a bit more real to everyone. But still, I didn’t feel that my colleagues (apart from the PhD students I spoke to) realised what they could do themselves on their computers.
Some colleagues say that this will only be transitory, that the gains from using LLMs will quickly be outweighed by the costs. I think the opposite. The first time I was impressed by an LLM was with DeepSeek in January 2025. I gave it a pretty hard undergraduate analysis problem, and it managed to solve it. This wasn’t worrying at the time. My point is that DeepSeek needed data center grade hardware. Now models that can solve this kind of problem, even faster than DeepSeek, can run on my computer (a high-end consumer computer, but still). My guess is that models will become far better than they are now, and at a fraction of the cost (in any units you like).
I have a few ongoing projects, and I am pretty sure they could nearly all be solved in one shot in half an hour with the best version of Astra. I could do that, polish the result, send it off, and produce an article in a month that would not have to fear the comparison to my earlier productions. I believe people are already doing this[1]: I was looking at math.PR on Monday and thought, “That’s a lot of single-author articles today!” So I tasked my favourite agent with investigating how the number of single-author articles had changed, and I was not surprised by the result[2].
(Be re-assured: no humans were harmed in the production of these plots. In particular, none had to work or even think—most importantly, not me…)
I could finish every single project I have in a few weeks. I won’t do it because I still have some affection for these problems. But once they’re settled, I don’t think I will ever write an article the way I did before.
So what should we do?
First and foremost, I really think the most important step we should take as a community is to get rid of the idea of “ownership” of mathematical results altogether. It doesn’t make any sense to me now. Is it your theorem if AI proved it? Is it your theorem just because you did it first, if I could obtain the same result in ten minutes with Astra? To be honest, the problem existed well before: how many people have been listed as authors just for asking the question? That doesn’t feel so different to me from prompting an LLM.
It is a hard step, because it is so ingrained in us. Our whole professional life is built on the results we have fathered. Our whole idea of ourselves as mathematicians is constructed around it. All our myths, all our culture. Maybe I am not well placed to say this, because I’ve always been a mediocre mathematician, but look at the dispute over Navier–Stokes, fighting over who prompted the LLM better or first [4]. Who will be king of the ashes? I think it doesn’t matter anymore and, most importantly, it shouldn’t matter to us.
One could argue that disclosing AI use can solve the problem of ownership, but it is not verifiable. And I don’t trust people in a competitive environment. Anyway, the culture of priority in mathematics has always been toxic. If we get rid of it, I believe we can all be happier.
Concerning the practice of mathematics, I’d say, “you do you”. We don’t own mathematics, we don’t own the problems. People shouldn’t be ashamed of using AI.
On the other hand, if you find meaning in doing everything by yourself, good. It is just not the way I am built. I see no sense in spending months working on something that could be so easily solved by chatGPT. I can’t accept putting my name on an article that relies substantially on AI for doing the maths, either. That makes no sense to me. Maybe I will try to do mathematics in the following way. A colleague recently talked to me about the question of the large particle number limit of a two-dimensional log gas confined to a disk. The result is known only in a very particular case where the process has a determinantal structure. I think I will create a GitHub repository containing the results known either from the existing literature or through AI. Anyone will be welcome to contribute in any way they want. The goal would be to produce a coherent, digested account of the mathematics which, once it covers the question sufficiently, could be put on arXiv in the name of a team.
I would like to end with a word for PhD students and postdocs. If you, reader, are lucky enough to hold a permanent position, please be aware that they are probably far more aware of this situation than you are. In particular, if you disagree with me and think we are not going through a terrible transition, please acknowledge at least that they have good reason to feel scared and lost. Please allow them to use AI in any way they want, or not at all if they don’t want to. And more importantly, if you are involved in recruitment of any kind: stop counting candidates’ articles. Whatever people say, this has always been a very important criterion. It was already a bad one before, now it is just terrible. It creates a very wrong incentive. They need to hear that they don’t have to use AI to remain “competitive” and the community must acknowledge it.
At least to some extent
This is off-course not a proof of anything but we should not be naive.
Between August 2025 and August 2026 the number of single authored articles went from 1,174 to 2,603. Among the 2,603 articles of August 2026, only 809 mentioned AI (take these numbers with caution, they are uncertain)
One could argue that there is still human input in one of the teams, but how do you even measure it? It seems they used AI quite extensively.