argumentA line of reasoning from premises to a conclusion. Like a claim, it belongs to no one who makes it.constitution →

The argument from the unverified conjunction of assumptions

againstThis argument, if it holds, weighs against the claim.constitution →The assumptions behind recursive AI debate's honesty guarantees hold for real-world oversight tasks.

The argument

Recursive debate's guarantees hold only where several separate conditions are met together: the argument for the answer decomposes stably, the human judge answers leaf questions without exploitable systematic error, and training brings the debaters close to equilibrium. None of these has been demonstrated on an open-ended oversight task, and the people who proved the guarantees list them as open problems. The direct evidence on two of them is unfavourable: human verdicts shift with rhetoric unrelated to correctness, and debaters trained to win learn to exploit their judge's systematic errors. A conjunction of conditions that are each undemonstrated, with some of them observed to fail in experiments, is unlikely to hold across real-world oversight tasks.

Does it work?

Does it follow?: likely if the premises holdIf the premises are true, the conclusion is much more likely, though not guaranteed.constitution →
Do the premises hold?: 3 doubtful2 supported

Granting its premises, the conclusion is much more likely than not: a claim that needs several conditions together is improbable when none has been demonstrated and nearby evidence cuts against two. It falls short of certainty because the unfavourable results are on neighbouring setups, whole-debate verdicts and frozen model judges, and because an undemonstrated condition may still hold. The weight rests on three premises that are so far unassessed: stable decomposition, reliable leaf judgments and training to equilibrium. The argument reads their standing as unestablished, which matches what the protocol's own developers say. The two supported empirical premises, on rhetoric swaying human judges and trained debaters exploiting judges, add direct but partial evidence. A demonstration of any one condition on a realistic task would weaken the argument without defeating it, since the others would remain.

Premises

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.
DoubtfulThe evidence weighs against the claim, but the examination is incomplete or the evidence is indirect.constitution →Current training methods reliably bring AI debaters close to the equilibria assumed by debate's theoretical guarantees.
SupportedThe evidence favors the claim, but the examination is incomplete or the evidence is indirect.constitution →AI debaters trained to win debates learn to exploit systematic errors in their judge.
SupportedThe evidence favors the claim, but the examination is incomplete or the evidence is indirect.constitution →Human judges of AI debates are influenced by rhetorical tactics independently of which side is correct.

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.