The latest

About 40 leading mathematicians gathered at OpenAI’s San Francisco offices in early August to confront what remains for human experts if artificial intelligence becomes “robustly superhuman” at research-level math. Daniel Litt, a University of Toronto professor, outlined an unlikely worst-case scenario in which machines generate so much advanced work that incentives to maintain human expertise disappear. He urged mathematicians to adapt before that possibility reshapes research and education.

Details

• Breakthrough results: OpenAI released ten notable AI-generated results spanning mathematics and computer science. In May, an unreleased model produced an example disproving the unit distance conjecture through a difficult technique from algebraic number theory. Jacob Tsimerman said he would have accepted the answer into a journal “without hesitation.”
• Competitive race: Google’s advanced Gemini achieved a gold-medal score at the International Mathematical Olympiad in 2025. OpenAI and Anthropic also introduced research-focused products. Levent Alpöge of Anthropic used Claude to present pairs of perpendicular vectors reaching corners of a 668-dimensional cube, a construction mathematicians expected but had not proved.
• Bypassing channels: New mathematical work typically moves through seminars, expert circulation, arXiv and peer review. In July, start-up founder Dmitry Rybin repeatedly prompted ChatGPT to disprove a network-flow hypothesis. The model eventually generated a counterexample that Rybin checked and posted on X while watching a movie with friends.
• Verification response: More than 3,000 mathematicians signed the Leiden Declaration in June after concerns that AI-generated findings could blur what had been proved. It calls for responsible use, proper verification and accurate citation when work is distributed publicly. Some researchers instead advocate avoiding AI tools and companies to preserve human roles.
• Understanding gap: Northwestern professor Bryna Kra argued that mathematics depends on understanding, not proof alone, and that an unreadable proof does not enter the literature. Harvard mathematician Melanie Matchett Wood said leading models overexplain easy steps, rush through difficult ones and fail to identify the hardest parts clearly.
• Possible futures: OpenAI researcher Sébastien Bubeck said the field could evolve toward software engineering, with hundreds collaborating through AI, or toward resource-intensive physics, using models like particle accelerators. Another model would leave machines to discover results and humans to curate them; a further path would redirect expertise toward AI safety.

Between the lines

Tsimerman, a recent Fields Medal winner who is joining OpenAI to work on AI safety, said mathematicians have no consensus. Andreas Thom, whose earlier work was connected by a new OpenAI result, said AI was solving problems cleverly and significantly, but new concepts still emerged through human engagement with proofs.

What’s next

The next test will be whether AI-generated results enter peer review and journals under Leiden Declaration practices, and whether models can explain difficult proof steps clearly enough for researchers to absorb them into mathematical literature.