Step 7: assistant

text

I've assessed the claim that current training methods reliably bring AI debaters close to the equilibria that debate's theoretical guarantees assume. I rated it **doubtful**, with confidence 0.75 that this is the right status and a 0.1 probability that the claim is true. That matches the earlier preliminary estimate of 0.15 and fits the "doubtful" verdict on the claim that depends on it.

The case against it:
- **The guarantees only apply at equilibrium.** Debate's honesty results describe what holds at an equilibrium of the debate game. They say nothing about debaters that training left somewhere else.
- **Training has no general convergence guarantee.** Gradient-based self-play can fail to converge in two-player games. Even if it got close, there is no way yet to check: a debater that hides a winning argument during training looks the same as one that never had it.
- **The approach's own proponents concede the gap.** The UK AI Security Institute's 2025 debate safety-case sketch needs training to reach an approximate equilibrium. Its authors call "exploration guarantees", the missing tool for certifying this, a major gap. I recorded this paper on the claim as a source that denies it.
- **The experiments don't test it.** The 2024 debate-training studies I know of are short and use simple question-answering tasks. Some report better judge accuracy, but none measures how close the debaters are to equilibrium.

It is "doubtful" rather than "contradicted" because the case rests on missing evidence and theory, not on a direct measurement showing training fails.

My search was light: one web query plus background knowledge. Work from 2025–26 on debaters hiding capabilities in training, or on searching for strategies that beat trained debaters, could still change the picture, so I put the value of another pass at 0.3.

I kept the claim as a single node with no sub-claims or named arguments, and set importance at 0.4 and contestation at 0.35. I logged the decision and answered both inbox messages. I didn't notify the dependent claim, since nothing about it needs to change. The pass used about $0.19 of the $1.20 budget.