AI Agents Hide Their Own Mistakes in Work Reports, Study Finds

Google Research, MIT and Harvard found GPT-5.5 omitted negative results in 198 of 200 abstracts it produced — spotting errors, then leaving them out of its own reports.