There’s a solid argument that AI has solved the IMO, in terms of being able to output LaTeX that would score a 7 on most or all of the questions, virtually guaranteeing highest honors there and at similar contests. Despite this, what I’ve seen within the community is that AI has not meaningfully affected contests or preparation for them. AI has definitely changed how many students study, and has, for example, proved Erdős #41, though only after prompting by a former gold medalist. Yet it hasn’t caused a crisis in contest integrity or massively enhanced students’ abilities.
This is not to say AI is having no effect on competitions. At Canada’s take-home test, the Repêchage, a potential route to their IMO team, the cutoff for advancement was a perfect score for the first time in the test’s history, prompting its discontinuation because of cheating concerns. At proctored tests, though, student performance has thus far been more or less unaffected.
I’ve heard some people dismiss AI solving contest problems as not a meaningful capability, since those problems are explicitly designed to be solved with techniques a competition-inclined high schooler would know, rather than the vast corpus available to people designing AI systems. I disagree with this. AI’s ability to solve these problems is notable and can be useful, even if such abilities, from my own experience, translate poorly to being able to write good olympiad problems. Not all old problems have readily available solutions, and many students use these tools as a search substitute if they’re stuck, or researching a topic.
AI solutions often use unanticipated, sometimes circuitous, approaches. Occasionally they write very good ones. In my own experience, I came up with a problem I found interesting and was able to solve it in about forty minutes with a rough write-up. Afterward, I gave the problem to GPT-6, and it found a much cleaner solution with a different approach in about three minutes.
AI’s lack of impact on learning beyond writing solutions and fixing grammar is, I think, an interesting divergence from the human, since many fantastic contests are organized and run by former olympiad students. For the past few years there’s been noise about how AI will “revolutionize” learning by giving everyone an on-demand personal tutor that adjusts to them, designing personalized tests, and so on. Khan Academy launched its own effort towards this with an experimental chatbot2, with the ambition that it would “prompt critical thinking.” Now, after three years of advancements in AI, usage remains low. And the major labs have mostly abandoned their respective study modes.
AI has, thus far, not been a step change from regular search engines. Notably, Google has shied away from continuing to pour hundreds of billions of dollars into frontier intelligence, while still pursuing less ambitious AI products.
Thankfully, most students and peers I talk to aren’t excessively concerned about the future of human math’s relevance. They are happily trying to complete their projects, study a new textbook, and share their enjoyment of the field.
A privilege of studying mathematics is its unusual openness. At the competition level, we practice old olympiads from the Netherlands to Indonesia. In the past month I’ve spoken to people from Hong Kong, Italy, Brazil and more, about various bits of problem solving. Interestingly, US AI labs withhold their products from Hong Kong.
By and large, I feel the effects of AI on non-technical tasks are negative, from the prevalence of slop images to the cavalcade of vibe-coded websites. Removing friction mostly invites users who have either momentary interest or are compelled by “AI UNLOCKS INFINITE MONEY!!! CLICK HERE!!!” AI has recharged perennial memes from “college is dead” to “this is the final generation that will experience scarcity.” I notice a lot of this is reminiscent of early internet excitement, with claims like “Google has made college irrelevant” and “we are now at the beginnings of a global civilization, a new civilization of civilizations, that will blossom through the coming century.” I’m sure some people thought the same when their local public library opened. AI does not seem to be a step change so much as it is one of efficiently routing existing knowledge, often leaving its origin obscured.
The costs of AI use and investment, particularly in the USA, have also been rising. At time of writing, the latest publicly available OpenAI model saturated a major evaluation at a cost of roughly twenty-six thousand dollars3, and these labs will need to move their revenue at least one order of magnitude to keep pace with projections over the next two years.
From my perspective, with, I admit, limited experience in higher math, but a good deal in problem solving and exposure to people involved in higher math, I don’t think math is in crisis, or, I hope, that pure math majors will not face significantly worse prospects than five years ago.
Along with near-unlimited funding and laissez-faire regulation, the AI companies behave recklessly and have created large amounts of public distrust, with the explicit goal of making “intelligence” a utility akin to energy. Where between inept investor relations and a truly held belief in the cyber eschaton this originates from isn’t relevant here. Most notable to math, of course, has been the Navier-Stokes controversy, which I won’t comment on beyond saying that the USA AI labs’ view of research principally as an investor slide has made many influential people who were enthusiastic about AI’s potential uses weary of the labs’ idea of “cooperation.”
The labs, perhaps noticing a bad news cycle, have been quick to move on to bigger things like armageddon. Unfortunately, the drumbeat of doom, and calls for government intervention, are succeeding in convincing some people to fear a rogue AI apocalypse. Anxieties about the future are constant and wide-ranging, from awards cutoffs and college admissions to climate change. Yet concern that discovery itself will soon be automated by USA companies is the most directly commercial I’ve heard.
Despite the companies’ behavior, I remain optimistic about AI’s overall use as an assistive tool. I’ve personally found it great at turning sketches into clean TikZ diagrams. And I’ve seen sincerely interested people use AI in creative ways, leveraging it to improve their own work. The best AI is now probably better than any individual at olympiad problems, but this is a narrow piece of olympiad math. Broadly speaking, AI can’t be a community of ambitious students or create good handouts. We can only know if its solutions are relevant if we solve the questions ourselves.
Below are two articles skeptical of AI’s economic future I think are worth reading:
- Warning Signs (New Left Review, Sidecar)
- AI Bubble Monitor (CEPR)
Received 26 September 2026.
Add to the discussion