AI systems produced valid, novel solutions to many open mathematical problems in 2026.
Assessed first pass
On a tight budget, assessed as supported (credence 0.85) from web evidence: Erdős problems verified by Tao, the unit-distance counterexample, the August Astra results, and the October 722-manuscript release with its withdrawals. Left atomic with no decomposition. A future pass could add subclaims on the validity rate of the October corpus and on the ETH refutation. Recorded one instance (Quanta).
Assessed Supported
verdict confidence 0.75 · credence 0.85
The evidence favors the claim. Through 2026, a run of AI-produced results on open problems was checked by independent mathematicians: several Erdős problems in January 2026 whose proofs Terence Tao reviewed and catalogued (Tao also cautioned that many were relatively accessible "low-hanging fruit"); an OpenAI model's May 2026 counterexample to Erdős's 1946 unit distance conjecture, which Quanta Magazine described as the first historically significant proof to come from an AI model; and further Erdős problems announced in August 2026. Taken together, these independently checked cases already make "many" a reasonable description. The large claims from late 2026 are less settled. In October 2026 OpenAI released 722 manuscripts in 372 result families from an unreleased model. Three were withdrawn within a day and others revised, and outside verification is still under way, hampered by the lack of access to the model and prompts. OpenAI's late-October claim that the Exponential Time Hypothesis is false comes with a Lean verification but is only weeks old. The claim does not depend on these, but how large "many" turns out to be does. Two things remain open: how many of the October manuscripts survive expert checking, and how far "novel" holds up against rediscoveries of results already in the literature, which affected some early Erdős-problem claims. Because "many" is vague, the claim carries a little definitional slack.
Claim entered the graph