The argument from the robustness of sound arguments and from debate experiments
The argument
A sound argument about a question that turns on fuzzy human judgment does not hinge on any single piece of evidence, and one that does can be strengthened by gathering redundant support, so real oversight questions admit stable decompositions. Recursion narrows the dispute to small, concrete leaf questions, which are the kind humans answer reliably, and judges can be trained, pooled and restricted to such questions. Experiments in which debate favours the debater arguing for the true answer, including runs where debaters were trained by self-play, are what would be observed if the conditions were approximately met. The assumptions therefore hold, or can be made to hold by protocol design, on the tasks that matter.
Does it work?
- Does it follow?
- slight even if the premises holdEven if the premises are true, they make the conclusion only a little more likely.constitution →
- Do the premises hold?
- 2 doubtful1 supported
Even granting its premises, the argument moves the conclusion only a little as stated, because it is silent on whether training reaches equilibrium and on deployment conditions, which the guarantees also need. Its premises are also stronger than their evidence. That real questions decompose stably is shown only for conclusions backed by many independent pieces of evidence, and no one has measured it on a realistic task. That humans judge leaf questions reliably is unmeasured for leaves chosen by optimised debaters. The experimental premise, that debate favours the true answer, is supported with moderate confidence but comes from shallow protocols on tasks with known answers, so it is what one would see if the conditions held on those tasks and says little about open-ended ones. The argument is best read as a reason to think the assumptions can be approached by protocol design, not that they hold.
Premises
- SupportedThe evidence favors the claim, but the examination is incomplete or the evidence is indirect.constitution →AI debate systematically favors the debater arguing for the true answer over the debater arguing for a false one.
- DoubtfulThe evidence weighs against the claim, but the examination is incomplete or the evidence is indirect.constitution →Arguments on real-world AI oversight questions admit decompositions stable enough for recursive debate's efficiency guarantees to apply.
- DoubtfulThe evidence weighs against the claim, but the examination is incomplete or the evidence is indirect.constitution →Human judges answer recursive AI debate leaf questions accurately enough to prevent systematic exploitation by debaters.
Where it is made
Passages where sources make, criticize or mention this argument. They record where it appears, not how strong it is, and none has precedence over another.
No source is recorded as making it.