The Claim and the Counterclaim

In a paper titled "OpenAI's claimed disproof of Connes' Rigidity Conjecture is invalid," Héctor Zenil and co-authors have systematically dismantled the proof that OpenAI's o1 model allegedly produced for the Connes rigidity conjecture. The paper, available on PhilArchive (https://philarchive.org/archive/NIEWTCv17), argues that the proof contains critical logical errors, making it invalid.

This is not a trivial academic dispute. The Connes rigidity conjecture is a deep result in operator algebras, with implications for von Neumann algebras and ergodic theory. If true, it would classify certain group actions. If false, it opens new avenues. The stakes are high, and the AI-generated proof was initially met with excitement. But Zenil's team shows that the proof's foundation is shaky.

Specific Flaws Identified

The critique points to several concrete errors:

  1. Misapplication of theorems: The proof incorrectly uses results from Connes' earlier work, ignoring hypotheses. For instance, it applies a theorem that requires the group to be ICC (infinite conjugacy class) but the proof doesn't verify this condition.
  2. Unjustified steps: A key step assumes that a certain map is an automorphism, but this is not proven. The authors show that the map fails to be well-defined under the given conditions.
  3. Circular reasoning: The proof uses the rigidity conjecture itself to prove a lemma that is then used to prove the conjecture. This is a classic logical fallacy.

Technical Details

The paper provides a line-by-line analysis of the proof. For example, it identifies that in the proof, the authors of the original claim define a group action on a probability space and then assert that the action is free. However, Zenil et al. construct a counterexample where the action has fixed points, contradicting the assertion.

Another issue: the proof relies on a result about the fundamental group of a II_1 factor, but the proof of that result is itself flawed. The critique points to a specific equation (Equation 7 in the original) where a limit is exchanged without justification.

The paper also notes that the proof does not address the case where the group is not amenable, which is central to the rigidity conjecture.

Implications for AI-Generated Mathematics

This incident raises serious questions about the reliability of AI-generated proofs. The OpenAI model produced a plausible-sounding proof, but it failed under scrutiny. This is not the first time AI has generated flawed mathematical arguments, but it is one of the most high-profile cases.

For developers and researchers, this is a cautionary tale. AI tools like OpenAI's o1 can assist in research, but they cannot replace human verification. The paper's authors argue that "the proof is not only invalid but also misleading, as it gives a false sense of progress."

What Should Developers Do?

If you're using AI for technical work, treat its output as a draft, not a final answer. Always verify with independent sources. For mathematical proofs, run them through formal proof assistants like Coq or Lean, which can catch logical errors.

For example, you could use Lean to formalize the steps of a proof. Here's a simple Lean snippet to check a basic theorem:

theorem add_comm (a b : ℕ) : a + b = b + a := by
  induction b with
  | zero => simp
  | succ b ih => simp [add_succ, ih]

But for complex conjectures, formalization is a massive undertaking. The point is: don't trust AI outputs blindly.

The Bottom Line

The Connes rigidity conjecture remains open. OpenAI's claimed disproof is invalid. The mathematical community has spoken, and the proof is rejected.

This is a win for rigorous science. It shows that peer review still works, even in the age of AI. But it also highlights the need for better verification tools.

Next Steps

Read the full critique at https://philarchive.org/archive/NIEWTCv17. If you're working on AI-generated proofs, consider formal verification. And if you're a mathematician, this is a call to action: the conjecture is still open, and AI won't solve it alone.

Stay skeptical. Verify everything.