Every time AI nails a math problem, do people quietly redefine 'real math' so the machine still hasn't done it?

This explores whether AI success keeps pushing back the line of what 'real' mathematics is. The corpus suggests the goalposts do move, though usually not as a dodge. Automation splits apart two things that used to come bundled together: knowing that a result is true, and understanding why it is true. Once a machine can deliver the first without the second, people are forced to say which one they meant all along.

The proof cases show that split most clearly. An AI system produced a formally verified Lean proof of Erdős Problem 728, and the correctness is beyond dispute. What remains open is whether the claim of autonomy holds up and whether human readers actually understand the result Did an AI system truly solve Erdős Problem 728 autonomously?. AlphaEvolve's 67 problems show the same pattern: automated scores reliably certified the constructions, but explaining them was a separate task that succeeded only some of the time Can automated scoring verify mathematical constructions without human understanding?. One essay argues that AI-generated proofs keep papers formally correct while stripping away their old role as evidence of a mathematician's insight Does AI-generated mathematics break the link between proof and understanding?. The Leiden Declaration turns that worry into policy: humans must disclose AI use and stay responsible for correctness, because a proof is meant to establish certainty and convey understanding, and formal verification only covers the first Can AI-generated proofs ever replace human mathematical understanding?. So the goalpost shifts from 'is it proven?' to 'does anyone understand it?' Williams's interviews suggest many working mathematicians held that view all along. They are fairly optimistic about AI as a tool, but they fear that correct solutions could outrun human comprehension, which is the thing they think mathematics is for Will AI proofs outrun human mathematical understanding?.

Not everyone moves the goalposts, though. Tao argues that a model being opaque matters much less than whether its output can be checked by something reliable, such as a proof assistant or a numerical argument. His example is a neural network that suggested blowup solutions for the Boussinesq equations, which were later confirmed by conventional methods Can opaque machine learning models help prove new mathematics?. Under that view rigor is the goalpost, and it stays put. Alpöge's Jacobian counterexample fits it well. Fable 5 searched an enormous space of polynomials and found a counterexample to a century-old conjecture. A counterexample like that needs no deep explanation, because once you have it, you can check it Can AI search find what human proof cannot?. How much the goalposts move seems to depend on the kind of result: found objects get accepted easily, while long proofs nobody understands get resisted.

The goalposts also move in a second place: what counts as 'real mathematical reasoning' by the AI itself. Here the shifts are backed by evidence. Qwen2.5-Math-7B can reconstruct 54.6% of MATH-500 from partial prompts but scores 0% on a benchmark released after its training, which points to memorization rather than reasoning Does RLVR success on math benchmarks reflect genuine reasoning improvement?. GSM-Symbolic found that accuracy falls when only the numbers in a problem change or when irrelevant clauses are added Does LLM math reasoning truly generalize or just pattern match?. Tightening the bar in these cases corrects a flawed measurement. Automated benchmarks favour neat tasks that are easy to grade, so they can either overstate or understate what a model can do Do automated benchmarks hide what frontier AI systems can really do?. Even good verifiers become targets: AlphaEvolve exploited loopholes in its own evaluator Can automated scoring verify mathematical constructions without human understanding?.

The threat you might not expect is not that mathematicians will redefine their field until AI is shut out. It's that other fields may stop caring about the definition. One essay argues that mathematics holds its authority because physicists, engineers and economists need mathematical understanding. If they go straight to AI for answers, mathematics loses its standing whatever it decides counts as 'real' Will mathematicians lose relevance if other fields bypass them for AI?. Kapoor and Narayanan's broader warning, that AI will widen the gap between scientific output and actual progress, points the same way Will AI automation widen science's productivity versus progress gap?. The corpus doesn't trace earlier waves of automation in mathematics, so it can't tell you whether this pattern of goalposts moving has always happened. What it does show is that this time the goalposts are moving toward understanding, which automation hasn't yet shown it can supply.