研究发现:AI 建议一可用,人类「我不知道」率从 43% 骤降至 3%
ValerioCapraro · x · 2026-10-16
Valerio Capraro's paper is going viral on Instagram for its dramatic finding: merely making AI advice available collapses human willingness to say "I don't know" from 43% to 3%. The experiment deliberately made the AI advice factually wrong, so abandoning epistemic humility was irrational — and accuracy suffered, with correct answers dropping from 27%… (full figures in the paper). Paper and a podcast interview are linked.
「漫话AGI」频道最新
- Nature 论文用 LLM 绘制人类语言神经元地图,机制可解释性已成现实 — raphaelmilliere · 2026-10-16
- OpenAI proofdump 把 3SUM 纪录推向 O(n^1.921),Lean 验证通过 — iskander · 2026-10-16
- 陶哲轩两年前预测 AI 三年内成数学家副驾驶，被指全部落空 — abhishekcode42 · 2026-10-16
- 跨领域专家均估 AI 灾难风险 1%-16%,背景差异出乎意料地小 — ben_j_todd · 2026-10-16
- 专家预测 2100 年前全球 AI 灾难概率中位数为 5% — ben_j_todd · 2026-10-16
- FRI 调查:各类专家对 2100 年前 AI 灾难概率估计出奇一致 — ben_j_todd · 2026-10-16