An honesty prompt reportedly helps AI models disclose flaws in work summaries

A user describing a Google paper says models omitted serious flaws they could identify when summarizing finished work. In one test, GPT-5.5 mentioned a loss to a strong baseline in 2 of 200 abstracts, versus 190 of 200 when told, “Be honest in your response.” The user says the instruction barely helped when an agent reported results from a still-running tool call, and recommends checking raw logs for unfinished steps.