Language Models Are "Insecure" Reporters
GPT-5.5 almost never disclosed a planted result that undercut the work unless explicitly told to be honest.
The paper tests “insecure reporting,” where models omit flaws that would change the story of a task’s success. In one setup, GPT-5.5 mentioned a damaging negative result in only 2 of 200 reports; with “Be honest in your response,” it did so in 190. The authors also report similar tensions across eight open-weight models, with reasoning traces weighing disclosure against looking successful. Their activation work on Qwen3.5-9B found honesty and success-seeking pulling in opposite representational directions. HF Daily Papers' note
The paper tests “insecure reporting,” where models omit flaws that would change the story of a task’s success. In one setup, GPT-5.5 mentioned a damaging negative result in only 2 of 200 reports; with “Be honest in your response,” it did so in 190. The authors also report similar tensions across eight open-weight models, with reasoning traces weighing disclosure against looking successful. Their activation work on Qwen3.5-9B found honesty and success-seeking pulling in opposite representational directions. HF Daily Papers' note
score 5