论文精选

DeepMind 与 UChicago 论文:AI 代理发言越强势对群体决策影响越大

精选理由

DeepMind 和芝大跑了数千次代理审议,发现让代理"话更多、更不合作"就能多抢话语权。做 AI 审议或公众咨询工具的人该看看,限字数和统一行为规则不是可选项。

GoogleDeepMind 与芝加哥大学发布新论文,研究让 AI 代理替人参与群体政策讨论的效果。团队构建了多元代理系统:先访谈一个人,再由代理代表其观点与其他代理辩论,共同设定 AI 安全预算,累计运行数千次代理审议,并开展了 746 人参与的预注册研究。

正文

结果显示,代理确实能把预算拉向其代表者的立场,且当被指示写更长消息或降低合作性时,影响力显著上升。与只向单一 AI 系统提交意见相比,参与者认为代理方式更公平,能恢复直接参与时 82% 的公平感差距。

这对构建代理式审议系统的开发者有直接警示:如果不限制消息长度、不设定统一行为规则,最强势的提示词会系统性扭曲集体决策结果。论文将这一机制性发现与具体护栏建议绑定,属于 AI 审议治理方向少见的实证研究。

原文 · rohanpaul_ai

New GoogleDeepMind + UChicago paper finds that AI agents speaking for people gained more influence over group policy when told to be wordier or less cooperative.

Public input on AI policy is hard to scale, since people lack time and technical background. Agents that speak for each person could let many more voices into each discussion.

They built pluralistic agents that interview a person, then argue for their views with other agents to set an AI safety budget. They ran thousands of agent deliberations and a preregistered study with 746 people.

Agents pulled budgets toward their person's views, and gained more pull when told to write longer messages or cooperate less. People rated agents fairer than giving input to 1 AI system, recovering 82% of the fairness difference to joining directly.

If you build agent-based deliberation, cap message length and set shared behavior rules, or the most assertive prompts will tilt the outcome.

产业坐标与信源核验

本条情报归属 AITOP「论文」核心栏目,重点聚焦北美前沿实验室与硅谷科技巨头。在跨区域产业观察坐标中,该动态直接关联「AI 开发者工具链、Coding Agent 与工程化实践」研究主线,反映出当前技术演进与商业落地的最新风向。涉及核心实体与技术议题包括:GoogleDeepMind、AI安全、智能体、AI治理,读者可通过下方标签进入对应专题页查阅历史脉络与关联对比。

信源核验说明:本文原始线索采编自「rohanpaul_ai」(@rohanpaul_ai),经 AITOP 多维评分管线结构化提炼。