AI and the Era of Automated Counterexamples in Mathematics

AI is Accelerating the Discovery of Mathematical Counterexamples

Artificial intelligence is fundamentally shifting the landscape of mathematical research by automating the discovery and formalization of counterexamples. In 2026, advanced LLMs—specifically OpenAI's Sol and models from the Fable family—have successfully disproved several long-standing conjectures, demonstrating a capacity to find "low-hanging fruit" in the form of counterexamples that had eluded human mathematicians for decades.

The Resolution of Major Conjectures

Recent developments have shown that AI can not only propose counterexamples in natural language but also formalize them in interactive theorem provers like Lean to provide absolute verification.

The Erdős Unit Distance Conjecture

In May 2026, ChatGPT disproved the Erdős Unit Distance conjecture in discrete geometry. The proof relied on a profound theorem in number theory from the 1960s (Golod and Shafarevich). While the initial result was informal, it was rapidly autoformalized into Lean by Logical Intelligence and later fully formalized by Boris Alexeev using the Sol model. The latter effort involved generating 1.2 million lines of Lean code, effectively proving hard theorems in global class field theory as a byproduct.

Grothendieck's Group Scheme Question

In July 2026, the Sol model discovered a counterexample to a 60-year-old question posed by Alexander Grothendieck regarding whether every finite free group scheme of order $n$ is killed by $n$. The AI identified a group scheme of order 4 that was not killed by 4. This result was autoformalized by Fable in approximately 1,000 lines of Lean code and subsequently merged into the mathlib library.

The Jacobian Conjecture

One of the most significant breakthroughs occurred when Fable found a counterexample to the Jacobian Conjecture, a famous problem in algebraic geometry that had remained open for 100 years. The counterexample, involving polynomials in three variables of degree 7, was manually formalized by Paul Lezeau and submitted to DeepMind's Formal Conjectures repository, providing a definitive resolution to the century-old problem.

The Role of Formalization and Lean

Formalization is the critical bridge between AI-generated "hallucinations" and mathematical truth. Because Lean is a programming language, AI-generated code can be verified by a compiler, removing the need for human trust in the AI's reasoning process.

• Autoformalization: Tools from companies like Logos Research, Logical Intelligence, and OpenAI are now capable of translating natural language mathematics into Lean code.
• Verification: Once a conjecture is formalized as a Lean statement, checking an AI-generated proof or counterexample becomes a trivial computational task.
• Scale: The speed of AI development is unprecedented; one researcher reported completing a project involving 250,000 lines of Lean code in just two weeks using Sol and Fable.

Expert Perspectives and Industry Impact

The integration of AI into high-level mathematics has met with mixed reactions from the academic community.

Academic Friction and Adoption

Some faculty members have expressed skepticism, suggesting that if a counterexample is "easy" for an AI to find, the problem was not mathematically interesting. However, others argue that the ability to actually find these examples is what matters, regardless of the duration the problem remained open.

The New PhD Workflow

There is a growing divide between researchers using AI tools and those who are not. Access to high-end models (costing approximately $200/month) is becoming a prerequisite for competitive research. Institutions like Harvard have already begun providing free access to Fable for PhD students, post-docs, and faculty to accelerate research output.

Synthesis of Community Insights

Discussion among practitioners highlights several key themes regarding this transition:

"The prompt for the Jacobian conjecture was obviously not random. the search space is too big to just try all the combinations of 3 variable polynomials."

This suggests that while the AI performs the heavy lifting of search and formalization, human steering and prompt engineering remain essential for directing the AI toward the specific areas of the mathematical landscape where counterexamples are likely to exist.

Furthermore, the community notes that counterexamples are not merely "bug reports" but are essential for refining definitions and sharpening the boundaries of mathematical truth, often leading to deeper human understanding once the AI-generated example is analyzed.