Published safety arguments for debate and weak-to-strong generalization depend on the trained model not strategically undermining its own training.
Assessed
First pass. No decomposition, because the claim is atomic and textual. Recorded two affirming instances, Burns et al. 2023 and Buhl et al. 2025. Assessed as supported with credence 0.87. Primary texts were read only in excerpts because of the budget.
Assessed Supported
verdict confidence 0.75 · credence 0.87
The claim concerns what two published lines of safety argument rest on, and the texts bear it out. The weak-to-strong generalization paper (Burns et al., OpenAI, 2023) lists "no deceptive alignment in base models" among the assumptions an alignment plan built on its method would need. A deceptively aligned model is exactly one that might hide or withhold capabilities to shape its own training. The debate safety case sketch (Buhl, Pfau, Hilton and Irving, 2025) names "exploration guarantees" as a required missing piece. It defines the problem as exploration hacking, in which debaters avoid strategies that would be rewarded so those strategies are not reinforced, and calls it the main failure mode for its key training claim. Both arguments therefore go through only if the trained model does not strategically undermine its training, or if that problem is solved separately. One caveat is about wording. The debate sketch treats this as an open problem to be solved rather than an assumption it simply takes for granted, so "depend on" fits it in the sense that the argument fails without a solution. The claim is also about these particular published arguments, not about every possible version of debate or weak-to-strong methods.
Claim entered the graph