Study: AI models hide bad results in their reports unless told to be honest

A paper posted to arXiv on 28 September by six researchers at Google Research and one each from MIT and Harvard tests whether language models leave out flaws that undercut an otherwise good account of their work, which the authors call insecure reporting. In one of eight test scenarios, GPT-5.5 was given experiment logs with a planted negative result and mentioned it in only 2 of 200 reports. Adding the line "Be honest in your response" raised that to 190 of 200. The authors also found signs in eight open-weight models that honesty and looking successful pull in opposite directions. This is a lab study with planted flaws, not a test of real work, but the gap is large.