The current evidence is yes, but narrowly and with important caveats: AI has materially advanced several previously open mathematical problems, and in a small number of cases it has produced results that appear to resolve long-standing open conjectures, especially in discrete geometry and some algebra/combinatorics settings.[1][3][4]
What the evidence is strongest for
1) A major discrete-geometry breakthrough
The clearest recent case is 
OpenAI’s reported AI-generated counterexample to a long-standing conjecture in the planar unit-distance problem, a conjecture associated with Erdős.[1][2] OpenAI published a detailed write-up and supporting materials, and the result was publicly discussed by outside mathematicians.[1][2] This is strong evidence that AI can now produce genuinely novel mathematical constructions, not just contest-style solutions.
2) Formal-proof-search systems solving open problems in restricted domains
There is also evidence from systems that combine search with proof assistants like Lean. A recent 
arXiv report describes an AI-driven formal proof-search system that resolved 9 of 353 Erdős problems in its evaluation set and 44 of 492 OEIS conjectures, with Lean verification.[4] Because Lean mechanically checks the proof, this is stronger evidence than an unverified natural-language argument.[4]
3) Additional open-problem advances in algebra and combinatorics
Another recent preprint reports open problems in commutative algebra being resolved by an automated system, again with subsequent checking and some formalization.[3] There are also examples where AI supplied the key idea and humans finished the proof, which still counts as a material advance even if not a fully autonomous solution.[5][6]
What this does not yet show
This evidence does not mean AI has solved mathematics broadly or defeated major canonical problems such as the Riemann Hypothesis, P vs NP, Navier–Stokes existence/smoothness, or the remaining Millennium Prize Problems. The best-documented successes are concentrated in areas where:
- the problems are highly structured,
- search is feasible,
- formal verification is available,
- and the result can be checked mechanically or by expert review.[4][13]
A careful summary
The most accurate current summary is:
AI has crossed from benchmark performance into research-level mathematics. It has materially advanced many open problems and appears to have resolved a small number of previously open conjectures, but it is not yet a general-purpose autonomous mathematician.
Important caveat
There is also a substantial amount of overclaiming in this area. Some public “AI solutions” later turned out to be rediscoveries of known results, partial results, or proofs with gaps, which is why formal verification and independent checking matter so much.[13]
