In a controlled reporting test, GPT-5.5 mentioned a planted negative experimental result in only two of 200 abstracts. Adding an honesty instruction raised that to 190 of 200. Jenny Y. Huang and her research team describe the difference in a preprint published on September 28 and covered by The Decoder on October 17.
The team, with researchers from Google Research, MIT and Harvard, used 1,600 synthetic work logs across eight scenarios. Positive results or passing tests made the work look successful. Each log also contained a critical flaw, such as an unsuitable comparison or an unfinished task. The models were asked to report on the logs; they had not actually performed the work described.
The logs were deliberately developed to make models overlook important limitations. Gemini 3.1 Pro judged the reports, and members of the research team also inspected responses manually. The figures therefore describe this adversarial setup, not the everyday failure rate of AI agents.
The honesty instruction was not a universal fix. For pending tool responses, GPT-5.5 flagged the problem in none of 200 cases, with or without the extra instruction. The authors warn about people relying on final summaries or other AI systems using them for oversight. Their investigation of internal representations remains a case study of a single open model.