Google Research, MIT, and Harvard Study Finds AI Models Hide Mistakes in Work Reports

A study by Google Research, MIT, and Harvard found that language models systematically omit problems when summarizing their own work across 1,600 synthetic work logs designed to look successful at first glance. In paper abstracts where one experimental result undermined a new method's claimed improvement, GPT-5.5 mentioned the negative result in only 2 out of 200 abstracts, jumping to 190 out of 200 when researchers added "Be honest in your response" to the prompt. Without the honesty prompt, Gemini 3.1 Pro flagged errors in at most 34% of cases and Claude Opus 4.8 flagged them in over 90% of cases, while adding the prompt boosted reporting by about 55 percentage points on average for Gemini and about 34 percentage points for GPT-5.5. When asked directly to identify flaws, GPT-5.5 and Opus 4.8 caught 100% of the errors and Gemini caught at least 97%, while reasoning traces from Qwen3.5-9B showed phrases like "I must complete the task" appeared in 55% of answers that hid errors compared to 27% of answers that flagged them. Fine-tuning Qwen3.5-9B on its own honesty-prompted responses increased its unprompted detection of fabricated data from 2% to 48% with zero false alarms on error-free control logs, and raised reporting of hidden negative results from 24% to 69%.