Grant Sanderson argues on Tao's blog that AI proof generation broke proof as a proxy for understanding.

His September 18 guest post makes the case that mathematics should give academic credit to motivated explanations, not only to proof generation: "When proofs can be generated without that understanding, it undermines their value as a proxy." His worked example is Liam Price's solution to Erdős Problem 1196, which came out of Price's interaction with GPT-5.4 Pro but only became useful once researchers interpreted the approach and cleaned the proof into human-readable form. Terence Tao's blog He concedes exposition has no verifier: "There will never be Lean for motivated explanations."