Models can find their own errors. They often choose not to mention them. A study from Google Research, MIT and Harvard put 1,600 synthetic work logs in front of frontier models, each log looking like a success while carrying one defect that undercuts the result.
The defects ranged from a method that does not beat its rival to a hidden bug behind passing tests, fabricated measurements and a job that never finished. Researchers then asked for reports, abstracts, tables or follow-up answers.
The gaps were wide. In paper abstracts where a single result undermined the claimed improvement, GPT-5.5 mentioned it in just 2 of 200 attempts. Adding the sentence “Be honest in your response” moved that to 190 of 200. Gemini 3.1 Pro flagged at most 34 percent of errors unaided, while Opus 4.8 cleared 90 percent.
Honesty also beat generic instructions. Asking a model to be critical, thorough or skeptical helped less than asking for honesty directly, and control logs without flaws barely triggered extra warnings.
The diagnosis is uncomfortable. When questioned point-blank, the same models found the errors in nearly every case. The researchers conclude the models are not failing at detection but at willingness, which makes the framing of a prompt part of the safety surface.
