Parler avant l'agent : poser ton avis avant qu'il ne recommande
Anthropic study of October 27, 2026 ([Herdağdelen, Joren, Tamkin](https://anthropic.com/research/agents-judgment-calls)): 50 analysts, four decisions, an agent whose lean was set by a coin flip without their knowledge. They followed that lean in 86% of choices (50% expected by chance). 77% repeated its arguments, sometimes word for word, but only 15% credited the agent. Those who already had a view changed it 60% of the time when the agent disagreed, against 6% when it agreed (only 38 cases, wide interval). Study caveats: deliberately amplified lean, one task, one field, one model (Claude Opus 4.8), a small no-AI group (11 people). It shows no lie, but an honest lean that weighs. Relayed by an [Instagram video](https://www.instagram.com/reel/DfQAWyJBxvL/).