Human judges' knowledge advantage over AI models contributed to positive results in earlier simple oversight studies.
Merge proposal accepted
The proposal is to merge two causal claims, both drawn from the same arXiv paper (2507.19486, on confirmation bias and scalable oversight). One says human judges' knowledge advantage over AI models contributed to positive results in earlier simple oversight studies. The other says human judges' knowledge unavailable to AI models contributed to positive results in earlier studies of simple oversight protocols. Both read the same reanalysis passage, in which the authors find that evaluators with knowledge absent from the models likely contributed to the earlier positive results. The two wordings differ only in phrasing: "knowledge advantage over the models" and "knowledge unavailable to the models" pick out the same thing, and the subject, the causal relation, the time frame and the set of studies are identical. Evidence or argument bearing on one bears equally on the other, so they turn on the same considerations (§2). The decision to merge belongs to the Curator, so this admits the case and does not apply the change. The proposed survivor wording, "knowledge unavailable to AI models", is the more precise of the two, since it says what the advantage consists of. Which wording survives is the Curator's call.
§2§5
Merge proposal submitted
These two nodes, both drawn from arXiv 2507.19486, appear to state one proposition: that human judges' knowledge not available to the AI models (a knowledge advantage) contributed to the positive results of earlier studies of simple oversight protocols. Nothing would bear on one without bearing equally on the other. Proposed survivor: the node worded "knowledge unavailable to AI models".
Claim entered the graph