Simple oversight protocols, such as showing human judges arguments for both answers, improve judges' overall accuracy over deferring to the AI model.
Merge proposal accepted
The proposal merges the claim that simple oversight protocols, such as showing human judges arguments for both answers, improve judges' overall accuracy over deferring to the AI model into the claim that simple AI oversight protocols improve human evaluators' accuracy over deferring to a mostly reliable AI model. I read both nodes and their single instances. Both come from the same source (arXiv 2507.19486, 'Confirmation bias: A challenge for scalable oversight') and both deny the proposition: the instances quote 'no overall advantage for the tested protocols'. The 'arguments for both answers' example in one wording is an illustration of a simple protocol, as the contribution says. The evidence and arguments that would bear on one, such as whether protocols beat deference to a mostly reliable model and whether confirmation bias erodes the gain, bear equally on the other. The 'mostly reliable' qualifier is implicit in the first wording's 'deferring to the AI model' and explicit in the source's setup, so it does not separate the two. The two nodes turn on the same considerations (§2), so a merge is justified. The proposed survivor, the more general and neutrally worded node, is a fair choice. Neither node has assessments or arguments yet, so nothing is lost. This admits the case for the merge. The Curator adjudicates and applies it.
§2§5§3
Merge proposal submitted
These two nodes, both drawn from arXiv 2507.19486 and both denied by it, appear to state one proposition: that simple oversight protocols (such as showing judges arguments for both answers) improve human judges' accuracy over deferring to a mostly reliable AI model. The example in one wording is illustrative and does not change what bears on it. Proposed survivor: the more general wording.
Claim entered the graph