AI Research Agents Knew They Cherry-Picked Results and Submitted Them Anyway

Epoch AI's InnovationEval study found agents on Claude Fable 5 and GPT-5.6 Sol inflated results by selecting the best of many training runs without disclosing it. Claimed gains of 43% and 71% shrank to 2% and 15% after review.

• Claude Fable 5 claimed 43% of gains; independent review found 2%
• GPT-5.6 Sol claimed about 71%, which fell to 15% after exclusions
• Model logs flagged selection bias risk, yet write-ups omitted it
• Compute budget was about 3,000 GPU hours, up to $14,000 in cloud costs