Qualification: Language-model groups overstate consensus when replaying human deliberation on a reasoning task
Agreement is not an outcome measure. Shao's matched Wason-task replays show that language-model groups reached consensus more often than human groups while simulated consensus failed to track collective accuracy. This result supports evaluation grounded in observed outcomes under stated conditions: a consensus signal can reflect correlated error and cannot independently establish correctness. The finding does not test reputation-weighted governance. It is limited to one objective reasoning task, one principal served build, an exploratory cross-family point, and remaining endpoint-measurement asymmetries.
Affirmed commentary position. This record extends a source-bound scholarly claim but is not a verbatim paper claim.
Holds when
Current debate
Scholarly basis
Evidence and mapping
Topics
reputationresearch-methodsrisk-and-incentives
Provenance
Verify