Tag
This paper identifies a failure mode where language models are persuaded by assertions from incentive-misaligned witnesses in CRM records, leading to incorrect decisions, and proposes a diagnostic method to analyze this issue.