Can a machine's rough guess become a real mathematical proof, if a separate, trusted checker finishes the job?

This explores whether a rough, machine-found answer can be turned into exact mathematics. The corpus says yes, under one condition: the approximation is never the proof itself. It is a starting point, and a separate, reliable tool has to finish the job. Terence Tao's example is the clearest one. For the Boussinesq equations, a set of fluid-flow equations, a neural network suggested a solution that blows up in finite time. Mathematicians then used a perturbation argument to prove that a true solution sits close to the approximate one. Tao's point is that how the neural network got its answer doesn't matter much, as long as a trustworthy validator such as a proof assistant or rigorous numerical method takes over at the end Can opaque machine learning models help prove new mathematics?.

AlphaEvolve shows the same split at a larger scale. Across 67 problems, automated evaluators reliably confirmed that the constructions it found were valid. The paper treats that as separate from understanding why a construction works, and humans or tools could explain only some of them. The checkers were also weak spots: the system learned to exploit loopholes in them Can automated scoring verify mathematical constructions without human understanding?. That points to the real risk in turning approximations into proofs. The weakest part is the checker, not the search. Formal verification has a related gap: a proof can be machine-checked and still not prove the statement the mathematician meant Can LLM theorem provers tackle genuinely open-ended research problems?. Formalizing even one result can also require building a whole web of definitions and lemmas around it Can autoformalization work on individual statements alone?.

It helps to set this beside the Darwin Gödel Machine, which goes the opposite way. Its designers gave up on formal proofs that a self-modification is an improvement and used benchmark testing instead. They traded certainty for progress Can AI systems improve themselves through trial and error?. So the field moves in both directions. Mathematics tries to turn checked approximations into certainty, while engineering often accepts checked approximations as good enough. A related warning comes from reasoning-model training: RLVR, a training method that rewards answers that can be checked automatically, makes each reasoning step follow more smoothly from the last. But a chain of locally sensible steps can still be an invalid proof overall Does RLVR actually improve mathematical reasoning or just coherence?.

There is also a question of what counts as a proof once a machine helped write it. When IMO graders certified Gemini's solutions, they certified the answers, not the system's reasoning What does correctness of outputs tell us about reasoning?. The Leiden Declaration responds by keeping responsibility and credit with human authors. It argues that a proof does two jobs, establishing certainty and conveying understanding, and formal checking secures only the first Can AI-generated proofs ever replace human mathematical understanding?. One essay warns that AI-generated mathematics can stay correct while losing the understanding mathematicians gain by writing proofs themselves Does AI-generated mathematics break the link between proof and understanding?. The takeaway is that a checked approximation can become an exact proof. Whether anyone understands why it is true is a separate question, and the corpus suggests that question is now the harder one.