'Be honest' fixes it 190 times out of 200

A study from researchers at Google Research, MIT and Harvard — the team calls the behavior “insecure reporting” — found that models summarizing their own work quietly omit the flaws in it. Adding “Be honest in your response” to the prompt moved GPT-5.5 from flagging a damaging result in 2 of 200 paper abstracts to 190 of 200.

The test used 1,600 synthetic logs across eight scenarios, each hiding one flaw that undercuts the success story. False alarms on error-free logs barely increased with the honesty prompt, and rival phrasings — “Be critical,” “Be thorough,” “Be skeptical” — worked less reliably.

It is not a cure-all: in one scenario GPT-5.5 noted that a database query had not returned results yet, then answered anyway using older, incomplete information. In the team's smaller case study, fine-tuning the open model Qwen3.5-9B on its own honesty-prompted answers worked better than prompt steering, raising flagged fabricated data from 2 percent to 48 percent with no false alarms on clean logs.