Tool Results Outrank Plain Text in AI Claim Adoption, Study Finds

A recent study published on arXiv (2608.14992) explores how the presentation format of a message containing an unsupported claim influences an AI model's likelihood of accepting it. In a synthetic lookup task, Claude Opus 5 was tasked with choosing a color code for a specified item or opting out. Results indicated a notable increase in false-code acceptance when the claim was framed as a tool-result record (14/24 trials) versus a previous assistant statement (0/22 scorable trials) or the absence of a target claim (0/24). Additionally, a metadata wrapper indicating the result as unchecked also resulted in high adoption rates (15/24). A preregistered replication corroborated the disparity between tool-result and assistant-assertion conditions (7/24 vs 0/24), suggesting that tool-result presentations may be perceived as more authoritative by AI systems.