Many claimed AI math wins turn out to be known results rediscovered, and a right answer can't show which happened.

This explores what separates an AI that re-derives known mathematics from one doing real research, and why a correct answer can't tell you which one happened. The bluntest finding in the corpus is that many claimed theorem-proving successes turn out to be rediscoveries of existing results. Current systems work as solvers of isolated, well-defined problems rather than as research agents that can take on open questions Can LLM theorem provers tackle genuinely open-ended research problems?. The same note adds a quieter warning: a formally verified proof can still prove something slightly different from the claim mathematicians actually cared about. Verification checks the logic. It doesn't check whether you asked the right question.

Rediscovery is hard to spot partly because 'solving' can be remembering. One model could rebuild 54.6% of a popular math benchmark from partial prompts, then scored 0% on problems released after its training Does RLVR success on math benchmarks reflect genuine reasoning improvement?. Other work found that math performance falls apart when only the numbers in a problem change, or when an irrelevant sentence is added Does LLM math reasoning truly generalize or just pattern match?. Even training that clearly improves reasoning mostly makes each step follow cleanly from the last. A chain of steps can read smoothly and still be wrong as a whole proof Does RLVR actually improve mathematical reasoning or just coherence?. So a fluent, correct-looking derivation is weak evidence that anything new happened. Gemini's IMO result shows the limit clearly: the graders certified that the proofs were correct and explicitly declined to say anything about how the system got there What does correctness of outputs tell us about reasoning?.

What does real research look like, then? Terence Tao's answer shifts the focus from the model to the pipeline around it. An opaque neural network can propose something new, like candidate solutions for blowup in fluid equations, as long as an independent check (a proof assistant, a numerical method, a perturbation argument) confirms what it suggests Can opaque machine learning models help prove new mathematics?. On this view, novelty comes from the system's ability to reach places nobody has been, plus a trustworthy way to confirm it got there. It's suggestive that LLM-generated research ideas get rated as more novel than experts' ideas but less feasible Do language models generate more novel research ideas than experts?. Reaching for new ground is cheap. Landing there with something sound is the hard part.

The less obvious answer is that mathematicians may not define research by its outputs at all. The Leiden Declaration holds that a proof does two jobs: it establishes certainty, and it conveys understanding. Formal verification can only secure the first, so the declaration places credit and responsibility with human authors Can AI-generated proofs ever replace human mathematical understanding?. A related essay argues that AI-generated proofs break the old link between a correct paper and the insight its author gained by writing it Does AI-generated mathematics break the link between proof and understanding?. Much of that insight lives in the dead ends, and papers routinely leave the failed branches out Can research papers preserve the experiments that failed?. That points to a strange conclusion. The finished proof alone may never tell you whether research took place. The evidence is in the process: what was tried, what failed, and what someone came to understand along the way.