Terence Tao, widely regarded as one of the greatest living mathematicians, has published a pointed critique of how AI-driven problem solving is reshaping mathematics, right after OpenAI published more than 300 mathematical results produced by an unreleased internal model. Tao doesn’t name OpenAI, but the timing, and the scale of the release, make the target hard to miss. His central claim is that solving open problems at industrial scale is “harvesting” them in a way that leaves entire fields less fertile than before.
What OpenAI released
On October 6, OpenAI put out a huge batch of manuscripts spanning 372 result families across much of pure mathematics and theoretical computer science. The model was posed roughly 4,000 problems in total, and the average result used about three hours’ worth of ChatGPT Pro thinking. Lean formalizations cover many, but not all, of the results, which means verification remains an open question for a large share of the release.
The drop came after months of escalating claims. In August, OpenAI said its Astra model had cracked ten long-standing open problems, and in September it shared a Navier-Stokes solution from a model it described as significantly more capable than Astra.
Tao’s argument: “Math 1.0” versus what AI is doing now
Tao begins by describing how a breakthrough works in what he calls “Math 1.0.” A major proof of a famous conjecture sets off a chain of activity: the authors give talks and meet other experts, workshops get organized, collaborations form, follow-up problems are shared, and new people, junior and senior, are drawn into the field. Over time, the proof is digested, streamlined, put in context, and eventually lands in textbooks and lecture notes for the next generation.
AI tools, Tao says, can support all of those follow-up activities when used responsibly. But right now the opposite is often happening. Problems are being solved autonomously by “AI prompters” who lose interest in the broader field once their target is solved, and who often don’t understand the output well enough to answer questions, give talks, or engage with other mathematicians. The result, he says, is far fewer seminars, workshops and collaborations than a traditional breakthrough would generate, and few people joining the communities around these problems.
He also points to a chilling effect: promising open directions are now being withheld from the public because researchers fear being scooped by someone pointing an AI agent at them.
“Contamination” and unsustainable harvesting
Tao’s most striking point is that the damage is hard to reverse. A solved problem can’t be made unsolved again, and even knowing that a solution exists “contaminates” later efforts, by humans and AI alike, to find alternative routes that might reveal additional insights.
In Tao’s opinion, solutions to open problems are being harvested at large scale in an unsustainable way, leaving entire areas of mathematics much less fertile than when those problems were solved the traditional way.
Toward “Math 2.0”
Tao says the field’s old premium on being first to solve an open problem, even if the solution isn’t well understood, has now been optimized to the point of unsustainability. “Math 2.0,” he argues, will need to decenter raw problem solving and value progress more holistically: elevating exposition, community building, and the opening up of new directions of study. He believes AI can contribute to all of those too, but says it will take “more imagination and ambition” than simply aiming an AI agent at a list of open problems. The mathematical community, he adds, will also need to explicitly re-evaluate its criteria for education, publication and career advancement.
Part of a growing pushback
Tao’s post fits a pattern. On September 11, 25 Fields Medal winners, Tao among them, signed a declaration arguing that the way AI companies race to solve famous problems is damaging the discipline, and that the goals of the labs and of the mathematical community are “severely misaligned.” Tao has also been candid about his uneasy relationship with the industry, previously saying that OpenAI used only the optimistic snippets of a long conversation with him in one of its advertisements.
What makes his latest argument notable is that it isn’t a call to stop using AI in mathematics. It’s a claim that the current incentives, in which speed and volume of “solved” problems are the metric, are working against the very community and understanding that make those solutions valuable. For labs treating open problems as showcases for their models, it’s a pointed reminder that the problems were never just trophies.