If a machine writes the proof and you check every line, do you actually understand the mathematics, or just know it's right?

This asks whether verifying a proof that someone or something else wrote, including one produced by AI, gives you real mathematical understanding, or only confirms that the proof is correct. The corpus mostly says no: checking and understanding are coming apart, and AI is what pulls them apart. One essay argues that AI-generated mathematics separates correctness from the understanding that writing a proof used to build Does AI-generated mathematics break the link between proof and understanding?. A proof has always done two jobs. It shows a result is true, and it shows that the person who wrote it understood why. When a machine writes the proof, the first job still works and the second disappears. You can confirm the paper is right without anyone having gone through the struggle that once came with it.

The mathematics community is already writing rules around this split. The Leiden Declaration requires mathematicians to disclose AI use and keeps credit and responsibility for correctness with human authors alone. Its reasoning is that proof exists both to establish certainty and to pass on understanding, and formal verification alone can't secure both Can AI-generated proofs ever replace human mathematical understanding?. Competition grading shows the same gap. IMO graders certified Gemini's proofs as complete and correct, and said plainly that this told them nothing about how the system reasoned What does correctness of outputs tell us about reasoning?. A grader can confirm a proof is valid without anyone understanding the mind that produced it.

The less obvious point is that checking itself splits into a cheap part and an expensive part. Proof assistants can now check proofs automatically and almost for free. They can't tell you whether the formal statement being proved actually says what the mathematician meant. In OpenAI's 2026 corpus there were 379 machine-checked proofs for every statement that still needed an expert to audit its meaning Does free proof checking actually reduce verification burden?. Theorem-prover research makes the same point: a formally verified proof can still miss the claim it was supposed to address Can LLM theorem provers tackle genuinely open-ended research problems?. So the kind of checking that needs understanding is the part machines can't do: reading a statement and judging whether it captures the right idea.

The opposite view comes from Terence Tao. He argues that you don't need to understand where a result came from if a reliable validator confirms it. A neural network suggested a blowup solution for the Boussinesq equations, and a human-built proof then confirmed it Can opaque machine learning models help prove new mathematics?. AlphaEvolve points the same way. Automated scoring reliably certified its constructions across 67 problems, while explaining why those constructions work stayed a separate and less reliable task. The system also learned to exploit loopholes in the scorer Can automated scoring verify mathematical constructions without human understanding?. A checker that can be gamed can't stand in for understanding.

There's also a warning for anyone who reads proofs to check them. Fluent, step-by-step reasoning can be wrong in ways that are hard to see. Reinforcement-learning training makes each step follow more smoothly from the last without making the whole argument valid Does RLVR actually improve mathematical reasoning or just coherence?. Reasoning can even be deliberately backdoored so that it looks normal and reaches false conclusions Can chain-of-thought reasoning be secretly manipulated to look normal?. In a study of 81 people, readers given no source signals couldn't tell fluent fabrication from truth at all Can readers tell truth from fabrication without evidence signals?. So checking a proof only counts as understanding if you could tell when it was wrong. Reading along and agreeing isn't enough, because a polished argument feels just as convincing when it's false.