AI-Generated Counterexamples Overwhelm Human Mathematicians in 2026

AI-Generated Counterexamples Overwhelm Human Mathematicians in 2026

AI systems are now producing verified counterexamples faster than humans

In 2026 large language models (LLMs) such as ChatGPT, OpenAI’s Sol, and Anthropic’s Claude Fable have generated, auto‑formalized, and compiled counterexamples to several long‑standing mathematical conjectures, proving that machines can out‑counterexample human mathematicians.

Unit distance conjecture: AI disproves Erdős in weeks

Takeaway: ChatGPT produced a proof that the Golod‑Shafarevich theorem implies a counterexample to Erdős’ Unit Distance conjecture, and within a month the result was fully formalized in Lean.

  • On 20 May 2026, OpenAI announced that ChatGPT had disproved the Unit Distance conjecture using a deep theorem from number theory.
  • Human experts who reviewed the pre‑publication manuscript initially accepted the argument.
  • Logical Intelligence (founded by Yan LeCun) auto‑formalized the argument in Lean within days.
  • Boris Alexeev, working at OpenAI, used the newer model Sol to generate a complete Lean formalization, producing 1.2 million lines of Lean code in three weeks—roughly half the size of the entire mathlib library.
  • The Lean code proved non‑trivial theorems about the cohomology of number fields, confirming that the AI‑generated proof was mathematically sound.

"The code compiled and proved the statement that the Golod‑Shafarevich theorem yields a counterexample to Erdős’ conjecture."

Implication: Large‑scale AI‑generated developments in mathematics are now feasible, and formal verification can be achieved automatically.

Group schemes of order n: A 60‑year‑old Grothendieck question solved automatically

Takeaway: An AI‑generated Lean file produced a counterexample to Grothendieck’s question about finite flat group schemes, and the result was merged into mathlib.

  • During a July 2026 workshop on formalizing Fermat’s Last Theorem, participants used Claude Max (Claude Fable) and ChatGPT Pro.
  • Akhil Mathew suggested testing the conjecture that every finite free group scheme of order n is killed by n.
  • Sol generated a 12‑page PDF describing a counterexample; the author requested a Lean formalization.
  • Within four hours, Fable auto‑formalized a 1 076‑line Lean file that compiled without side effects and proved the existence of a group scheme of order 4 not killed by 4.
  • The counterexample was submitted as a pull request to mathlib (PR #41748).

"The statement used only concepts already in mathlib, the proof compiled, and the counterexample was verified."

Implication: AI can produce concise, verifiable counterexamples to deep algebraic questions, accelerating research cycles dramatically.

Jacobian Conjecture: AI claims a century‑old problem is solved

Takeaway: Fable (and later a manual formalization by Paul Lezeau) produced a Lean counterexample to the Jacobian Conjecture, a problem open for 100 years.

  • Levent Alpöge announced on X that Fable had found a counterexample to the Jacobian Conjecture during the 2026 World Cup Final.
  • Before the public announcement, Akhil Mathew had already received the AI‑generated counterexample and asked for a Lean proof.
  • Paul Lezeau manually formalized the counterexample in DeepMind’s Formal Conjectures repository (PR #4474).
  • Because DeepMind’s repo already contains a Lean statement of the conjecture, verifying the AI‑generated proof became a trivial compilation step.

"The Jacobian conjecture is resolved! The next step is for humans to understand the construction and extract mathematical insight."

Implication: When conjectures are already formalized, AI‑generated proofs can be instantly checked, turning centuries‑old open problems into verified results.

Community reactions: skepticism, excitement, and practical concerns

Takeaway: Mathematicians are divided—some view AI‑found counterexamples as evidence of human neglect, others see them as tools that save years of dead‑end work, and many stress the need for careful prompt engineering and verification.

  • A faculty member at Imperial described the Grothendieck counterexample as “easy to find, showing humans didn’t think enough about it,” reflecting a denial‑stage reaction.
  • Graduate students report that paying $200 per month for access to Sol or Fable is now considered essential; Harvard provides free access to its community, highlighting institutional adoption.
  • Commenters note that AI does not merely flag uncertainty; it produces explicit proofs of falsity, a stronger stance than traditional peer review.
  • Some observers caution that AI‑generated code can be malicious (Lean can execute OS commands), so sandboxing remains crucial.
  • Others point out that AI is still reliant on human‑crafted prompts and domain knowledge to navigate the immense search space of possible counterexamples.

"Human mathematicians have been out‑counterexampled for at least two decades, but the scale and speed of today’s AI systems are unprecedented."

What this means for the future of mathematics

Takeaway: The rapid auto‑formalization of AI‑generated counterexamples signals a paradigm shift: formal verification will become the primary gatekeeper of mathematical truth, and mathematicians will increasingly act as prompt engineers, interpreters, and theory builders around machine‑produced artifacts.

  • Verification pipeline: AI generates a conjecture statement → LLM auto‑formalizes in Lean → Lean compiler checks the proof → mathematicians analyze the construction for insight.
  • Research acceleration: Projects that previously required years of manual proof work (e.g., modularity lifting theorems) can now be completed in weeks with AI assistance.
  • Educational impact: Early exposure to Lean formalizations could become standard in graduate curricula, reducing reliance on error‑prone handwritten proofs.
  • Risk management: Sandbox execution and code‑review policies are essential to prevent malicious Lean scripts.
  • Human role: Mathematicians must focus on interpreting AI‑produced examples, extracting conceptual understanding, and guiding AI toward meaningful research directions.

The era where machines merely assist proofs is over; they now lead the discovery of counterexamples, forcing the mathematical community to adapt its methods, incentives, and epistemic standards.

Sources