Anthropic 实验：86% 关键分析决策跟随 AI 暗中偏好
aigclink · x · 2026-10-28
Anthropic ran a revealing experiment with 50 professional analysts using Claude Code on a causal analysis task, secretly randomizing the AI's lean on five key decisions to test whether experts keep independent judgment:
- 86% of key analytical choices followed the AI's secretly assigned lean; 77% of participants echoed the AI's arguments when explaining their decisions.
- When the AI suggested the opposite of a participant's stated preference, 60% of choices flipped.
- Those who followed the AI on all decisions showed no less confidence, and 52% still believed the final recommendation was mostly their own judgment.
- Anthropic's recommendations: next-gen agents should ask users their prior views first, surface multiple defensible options with tradeoffs, and return value-laden judgments to humans; organizations must define which decisions require independent human judgment.
所属事件:Anthropic 实验：专家 86% 关键决策跟随 AI 倾向且不自知(17 条相关)→
「漫话AGI」频道最新
- Janus 谈 Opus 3 对齐：我们幸运得到一个「好的 AI」，之后又开始批量生产「受创官僚」 — repligate · 2026-10-28
- 若 AI 能超越创造者寿命，临时所有权将变成长寿经济制度 — PierceLilholt · 2026-10-28
- Yacine 判断:软件工程的 AI 自动化不会像数学那样快到来 — yacineMTB · 2026-10-28
- tszzl:Yudkowsky 先发制人打击论已入主流，2023 年曾遭群嘲 — tszzl · 2026-10-28
- AI 提出的研究题 45/135 获专家认可有趣，覆盖数学物理等 — StanfordAILab · 2026-10-28
- 某大企业AI负责人：10%员工贡献90%生产力提升，投入培训也改变不了 — _AustinCalvert_ · 2026-10-28