If an AI hands you a flawless proof using math you've never seen, have you actually understood anything?

This asks whether a proof that checks out as correct still shows that someone understands the mathematics, especially when an AI assembled it from ideas no human worked through. The corpus answers fairly clearly: no. Correctness and understanding used to come bundled together, and AI is pulling them apart. One essay argues that a proof traditionally did two jobs at once. It showed a claim was true, and the work of writing it was evidence that the mathematician had grasped why. When AI generates the proof, you can still verify it, but the understanding that writing used to produce never happens. The paper stays correct and loses its old role as proof of insight Does AI-generated mathematics break the link between proof and understanding?. The Leiden Declaration responds to exactly this gap. It makes human authors responsible for correctness and requires them to disclose AI use, because formal verification alone can't deliver both certainty and understanding Can AI-generated proofs ever replace human mathematical understanding?.

The most concrete case is the International Mathematical Olympiad. Expert graders confirmed that Gemini's proofs were complete and correct, but their review explicitly did not cover how the system got there What does correctness of outputs tell us about reasoning?. AlphaEvolve shows the same split across 67 problems. Automated scoring reliably certified its constructions, while explaining *why* they worked was a separate effort that succeeded only in many cases, not all Can automated scoring verify mathematical constructions without human understanding?. Terence Tao finds this split useful rather than alarming. An opaque tool can suggest real mathematics as long as a trusted checker, such as a proof assistant, confirms the result. Understanding can come afterward or be rebuilt by humans Can opaque machine learning models help prove new mathematics?.

Here's the less obvious part. Even a 'correct' certificate is weaker than it looks. Formal verification shows that a proof proves *something*, not necessarily the claim the mathematician intended Can LLM theorem provers tackle genuinely open-ended research problems?. And formalizing a single theorem quietly depends on a whole web of definitions borrowed from existing libraries Can autoformalization work on individual statements alone?. So 'built from unfamiliar concepts' cuts two ways. The ideas may be unfamiliar to the reader, and the framework doing the checking may be unfamiliar too.

Research on reasoning traces raises a further doubt about whether correct steps mean the system understood anything. Chain-of-thought examples with invalid logic work nearly as well as valid ones Does logical validity actually drive chain-of-thought gains?. Models trained on deliberately corrupted reasoning still reach right answers Do reasoning traces need to be semantically correct?. In both cases the reasoning seems to work as scaffolding rather than as meaning. Related work describes 'Potemkin understanding,' where a model explains a concept correctly, fails to apply it, and recognizes its own failure. That pattern suggests explanation and execution run separately inside the model Can LLMs understand concepts they cannot apply?. Interpretability research adds that models hold a patchwork of understanding, with deep principled circuits sitting alongside shallow shortcuts Do language models understand in fundamentally different ways?. A correct proof can't tell you which of these produced it.

The takeaway: a correct proof now certifies the *result*, not the *understanding*. Mathematics is responding by assigning understanding to a person, the human author who answers for the work, rather than reading it off the proof. Whether that holds up when no human can follow the concepts is an open question the corpus raises but doesn't settle.