An LLM verifier rated math proofs near-perfect; an expert found 17% correct

Jun 12, 2026 · 6 min read Easier to Convince Than to Prove aimathematicsverification Contents Easier to Convince Than to Prove The paper, in its place The number Why it doesn’t escape Footnotes Easier to Convince Than to Prove Three weeks ago I ended a post on a sentence from Melanie Wood: “in many cases, it will be easier for AI to convince humans it has a proof than to come up with a correct mathematical argument, and I believe that we as mathematicians are not sufficiently prepared for this.” I called it the dark twin of the verification story. I’ve now spent two posts on it — once as Wood’s forecast, once as DeepMind’s failure analysis, the agent burying a problem’s difficulty inside a sorry or citing a lemma it had hallucinated. The gap between convincing and correct can be sized.

…

Excerpt limited to ~120 words for fair-use compliance. The full article is at Korbonits.