Claude's FLT Proof Ignites Math Community Scooping-Ethics Debate

On September 7, Anthropic's formalization of Fermat's Last Theorem (FLT) in Lean using Claude triggered a broad math-community debate over scooping norms, attribution, and the boundaries of AI participation in mathematical research. Opinions split: some argued Anthropic should have proactively collaborated with Kevin Buzzard, who has long worked on the direction; others countered that obvious research goals cannot be monopolized. Buzzard himself took a conciliatory tone, only requesting an errata list. The dispute is a landmark case for how AI labs should behave toward academic communities.

Confirmed

• The thread began with @nihilunbounded invoking the unwritten norm against scooping in-progress research and mocking Anthropic as caring less about mathematics than OpenAI. Some Buzzard followers derided the mass AI-generated Lean proof as "Leanslop" and "graceless."
• @littmath said most mathematicians in Anthropic's position would have offered collaboration with Buzzard; he condemned rushing results before careful verification, but clarified he only objects to scooping in-progress work, not the norm against writing up already-completed work. He also noted that normal human projects naturally evolve into collaboration, which this AI proof failed to foster.
• @jacobaustin132 revealed Claude followed Wiles's original proof rather than Buzzard's planned simplification, making it harder; the team is considering contributing Tau Ceti. He communicated with Buzzard on Lean Zulip, whose main request was an errata list—partially completed—and Buzzard is genuinely interested in the formalization. Terence Tao's team also engages with the Lean community.
• @nihilunbounded argued Buzzard's goal was grander than formalizing FLT; taking out the most obvious subgoal damages an entire research program, and scooping norms protect senior researchers too. Even after attending lectures, one shouldn't use AI to validate the strategy and claim the result first.
• @alzzyd took the opposite view: proving FLT in Lean is an obvious goal, and long-term investment creates no de facto monopoly; if Buzzard's goal was FLT itself, he "lost" and deserves no credit. He also clarified that "scoop" in science means independently obtaining a result first, not exploiting someone else's told-to-you idea.
• @jacobaustin132 warned AI proofs could split the Lean community between human-readable and AI-generated styles, envisioning a giant theorem-statement graph where anyone burns Claude/Codex tokens to prove nodes.
• Separately, @CsabaSzepesvari questioned reliability: the correctness chain—compiler, Lean kernel, Mathlib, axioms—must all hold, and he worries about whether the formalization faithfully captures the original proof.
• On September 9, @Afinetheorem and @JacquesThibs debated another AI-problem-solving controversy: JacquesThibs argued the other party could technically publish proof details and Sholto's statement was merely de-escalation; Afinetheorem countered that writing out the proof's workings would give it away, and subproof details could expose the proof path.

Unconfirmed

• The identity of the proof in question and whether its subproofs would truly expose the proving path remain speculation without verification in the posts.

Why it matters

• The debate extends academic scooping ethics into the AI-lab era: if LLMs can do mathematics at scale in parallel, the community's one-person-one-problem tacit understanding may collapse.
• It clarified that "scooping" spans distinct scenarios—in-progress preemption, write-up races, and AI validation of others' lecture ideas—each judged differently, directly informing future norms for AI in mathematics.
• The follow-up dispute over whether publishing proof details exposes the proof path signals a new recurring tension between transparency and competitive strategy in AI mathematics.
• Szepesvári's skepticism reminds us that even a completed formalization's trustworthiness depends on the reliability of the software infrastructure behind it.

2026-09-07 ~ 2026-09-09 · 26 related posts

• Math's unwritten no-scooping norm sparks debate over Anthropic vs OpenAI on math — alz_zyd_ · 2026-09-07
• Are hard math problems unsolved because only one person works on each? — alz_zyd_ · 2026-09-07
• Math's no-scooping norm: young researchers protected, big open questions fair game — nihilunbounded · 2026-09-07
• littmath: scooping finished-but-unpublished work is where the norm applies — littmath · 2026-09-07
• Counterpoint: you don't know the talk's outline is sound, so don't race to plug it — nihilunbounded · 2026-09-07
• Defining "scoop": independent priority vs. building on someone's outlined strategy — alz_zyd_ · 2026-09-07
• littmath weighs in: 'People should be sure before they announce' research results — littmath · 2026-09-07
• Debate continues: long prior work confers no dibs on an obvious research goal — alz_zyd_ · 2026-09-07
• Core of the dispute: Buzzard's goal went beyond formalizing FLT in Lean — nihilunbounded · 2026-09-07
• If Buzzard's goal was broader, he can keep working and get due credit, argues critic — alz_zyd_ · 2026-09-07
• FLT credit fight: if the goal was just proving it in Lean, Buzzard lost and gets no credit — nihilunbounded · 2026-09-07
• [source] Mathematicians weigh in on Claude's FLT proof: most would have collaborated with Buzzard — littmath · 2026-09-07
• Mathematicians debate: Anthropic should have offered Buzzard collaboration over FLT-in-Lean — alz_zyd_ · 2026-09-07
• littmath: Claude's FLT proof bypassed the natural collaboration a human project would have had — littmath · 2026-09-07
• Claude's FLT proof followed Wiles' original path, not Buzzard's simplified version — jacobaustin132 · 2026-09-07
• Terence Tao's team talks with Kevin Buzzard over Lean formalization — jacobaustin132 · 2026-09-07
• [source] Buzzard responds kindly to Claude's FLT proof, asking mainly for an errata list — jacobaustin132 · 2026-09-07
• Predicting a Lean community schism: human-readable proofs vs AI-generated slop — jacobaustin132 · 2026-09-07
• Anthropic accused of scooping mathematician Kevin Buzzard with AI-generated Lean proofs — teortaxesTex · 2026-09-07
• Csaba Szepesvari questions trust in Lean compilers and kernels for formal proofs — CsabaSzepesvari · 2026-09-07
• Szepesvari walks back Lean doubts: printed axioms plus sorry-free proofs suffice — CsabaSzepesvari · 2026-09-07
• RL researcher Csaba Szepesvari concedes Lean kernel bugs are real but FLT formalization likely stands — CsabaSzepesvari · 2026-09-07
• RL Researcher Csaba Szepesvári Questions Whether AI's Formalized Math Proofs Are Faithful — CsabaSzepesvari · 2026-09-07
• Csaba Szepesvari questions whether AI formalized proofs are faithful to what mathematicians mean — CsabaSzepesvari · 2026-09-07
• Proof drama escalates: releasing details would expose the method, argues critic — Afinetheorem · 2026-09-09

1 near-duplicate retellings: CsabaSzepesvari