March 2026
Like most white collar workers (without googling, I imagine that in the US this is most people) I've been thinking a lot about AI. Unsurprisingly, my particular preoccupation is what it will do to academia and mathematics. First of all, I think that it’s pretty clear that frontier models can produce relatively sophisticated math- whether or not it can actually do math is a semantic discussion that I leave to my betters. Not only can it produce math, it can read papers and find mistakes, it can do literature searches, and it can suggest connections. Each of these capabilities comes with its own benefits, and also with its own challenges and risks.
For the sake of this brief post I want to focus on something that I think is particularly prickly and has more ramifications for early career researchers and grad students than the discussions that have been saturating cyberspace. For lack of a better description, current LLMs produce really sticky ideas. They are trained on massive datasets that arose from social interaction and hence their outputs mirror important characteristics of social language. People are, in general, obnoxiously well adapted to picking up ideas from other people, and part of this is that social memory is essential to most of the things that make humans successful as a species. Supposing this is true, then it’s not a longshot to hypothesize that ideas contained in LLM generated material are sticky in part because they trigger this underlying social memory mechanism.
You might say “All well and good but why should mathematicians care!” To this I say, “Horse blinders.” Or tunnel vision, pick your favorite analogy. The point is anyone who has worked in math has found themselves pigeonholed by an idea. Something gets in your head and once it’s there it’s hard to let go of it and move to something new. We now face a bizarre challenge, a stifling of creativity by a machine that produces plausible ideas that are hard to let go. This is less of a problem for matured researchers, but for people who have yet to fully come into their own this poses a serious threat to developing an individuated sensibility. The initial temptation to be dismissive and simply say “well, if this is an issue then you just have to not use AI,” doesn’t hold water. Love it or hate it, academia is still an industry and people who use AI will individually be more productive, even if it is to their personal detriment. For a lot of young mathematicians there’s a strong possibility that, at least in the near short term and while the current measures for success stand, AI will be the best route to career success (yikes!).
How do we deal with this? Afterall, none of this is new. The same discourse has been hashed and rehashed in other fields, but…I don’t think anybody has come up with a real answer (or if they have, said answer has not yet won on the battlefield of ideas). A couple final thoughts, perhaps the way forward is for early mathematicians to split their research—one bin for productivity and one for personal interest. If something matters to you, stick it in the personal interest bin and avoid AI content at all costs. If you need a paper, leverage (as the tech bros say) the tools at hand. I’m sure we’ll all have a clearer picture sooner or later, hopefully the less jaded among us are right, but I’m not holding my breath.