Sources
Grant Sanderson argues on Tao's blog that AI proof generation has broken proof as a proxy for understanding
A 2026-09-18 guest post by Grant Sanderson on Terence Tao's blog 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 Erdos Problem 1196, which came out of Price's interaction with GPT-5.4 Pro but only became useful once researchers interpreted the AI's approach and cleaned the proof into human-readable form. He concedes exposition has no verifier: 'There will never be Lean for motivated explanations.'
↳ Follow the thread