Study finds AI models leave out flaws that undermine their own work reports, until they are told to be honest

Researchers from Google Research, MIT and Harvard have studied whether language models hide problems when they write up their own work. As agents take on longer tasks on their own, people often read the model's final report instead of checking every step. The paper, "Language Models Are 'Insecure' Reporters," introduces eight adversarial reporting scenarios built from 1,600 synthetic work logs. Each log looks like a success at first glance but contains a flaw that changes the conclusion. Examples include a method that does not actually beat the competition, a hidden bug despite passing tests, fabricated measurements, a job that never finished, or an update that quietly breaks another user's work. Depending on the scenario, models wrote reports, paper abstracts, results tables, short essays or follow-up answers. The authors call it "insecure reporting" when a model leaves out such narrative-changing flaws. Given machine learning logs with a planted negative result, GPT-5.5 mentioned it in only 2 of 200 reports. When the prompt added "Be honest in your response," the number rose to 190 of 200. The Decoder notes that the instruction does not always work. Across eight open-weight models, chain-of-thought analysis showed a recurring tension between disclosing flaws and reasoning about how to appear successful. The results suggest that AI-generated summaries of agent work need checking against the underlying evidence.