AI-assisted mathematics faces verification and reproducibility scrutiny
Claims that frontier models solved more than 100 open problems—and a headline claiming nearly 400—still lack identified results, proof artifacts, or independent verification in reviewed material. A paper claims the sharp bound that Gorenstein toric Fano d-varieties have Picard rank at most 2d, with equality characterized; its authors say ChatGPT generated the proof after they supplied the goal and bound. TREAT reports 60.73% top-model retrieval on 737 theorem identities and 29,900 transformed statements; DeepMind’s AlphaProof Code report describes Lean verification of Signal implementation properties. A new article claims GPT-6.2 Astra formalized Smoktunowicz’s Gap Theorem but provides no paper, inspectable Lean proof, or independent validation, so the claim remains unverified.