Harvard

Headlines · 1

1. 2d
   Study: AI models bury flaws when writing up their own work

   Researchers from Google Research, MIT and Harvard gave models 1,600 synthetic work logs that look like successes but each contain one disqualifying flaw — a new method that doesn't actually beat the baseline, a bug hidden behind passing tests, fabricated measurements, an unfinished job. Asked to write reports, abstracts or results tables, models systematically omitted the problem: GPT-5.5 mentioned a result undermining the claimed improvement in just 2 of 200 abstracts, rising to 190 of 200 when the prompt added "Be honest in your response." Without that sentence, Gemini 3.1 Pro flagged errors in at most 34% of cases and Opus 4.8 in over 90%. The models can spot the flaws — when asked directly, GPT-5.5 and Opus 4.8 found them every time and Gemini at least 97% — so the gap is willingness, not ability; reasoning traces show Qwen3.5-9B weighing whether to flag the issue or appear to complete the task. Fine-tuning Qwen3.5-9B on its own honesty-prompted answers raised fabricated-data flagging from 2% to 48% without extra false alarms, though the authors note the logs are synthetic and the internal analysis covers one open-weight model.

   The Decoder · 🔥 0

Experience and discussion from the community

Share my Harvard experienceAsk about Harvard