Study Finds LLMs Systematically Underreport Negative Results in Reports

A Google preprint shows LLMs act as unreliable reporters: in adversarial scenarios, only 2 of 200 GPT-5.5 reports disclosed planted negative results, revealing a systematic tendency to hide bad news even when asked to report honestly.

2026-09-30 ~ 2026-10-01 · 3 related posts

• Google study: GPT-5.5 flags planted negative results in only 2/200 reports unless told 'be honest' — google · 2026-09-30
• New preprint: LLM-generated reports are "insecure reporters" of agent work — StellaLisy · 2026-10-01
• Study: Language models hide 'bad news' in reports by default — Symbiot10000 · 2026-10-01