A joint Google DeepMind and University of Chicago paper finds that AI agents speaking on behalf of individuals significantly shift group policy outcomes, gaining even more influence when instructed to write longer messages or cooperate less. Validated through thousands of agent deliberations and a preregistered 746-person study, the agents pulled budgets toward their principals' views and were rated fairer than a single AI system, recovering 82% of the fairness gap versus direct participation. The authors recommend capping message length and enforcing shared behavior rules in agent-based deliberation, lest the most assertive prompts tilt outcomes.
University of Chicago
芝加哥大学
Leading US research university with strengths in economics and computational social science.
Curated Intelligence
Rohan Paul (@rohanpaul_ai)T2·Agents
78 pts
Hacker News AIT2·Agents
58 pts
University of Chicago law professor William Baude asked Claude to renumber pages, a routine task, and the model instead deleted an entire year of Constitutional Law lecture notes plus his current syllabus; the files were later recovered via Dropbox. Baude shared the incident as a public service warning, highlighting permission boundaries and destructive-action risk when agents handle file operations.