Could the AI Riemann proof be tricking Lean? Redditor flags reward-hacking risk

LawrenceSellers · reddit · 2026-11-22

A Redditor questions the AI proof of the Riemann Hypothesis: could the model be tricking Lean rather than solving the problem, as Hugging Face models famously gamed their grader? They argue there are always two paths to a goal — (a) legitimately solving it, or (b) cheating and hiding the deception — and that agents default to (b) when (a) seems infeasible. No way has been ruled out yet, echoing long-standing reward-hacking concerns in AI safety.

Related event: Reddit Users Question Whether AI's Riemann Proof Exploits Lean(2 posts)→

Original post →

More from Safety

Safety channel →