AI Trends

Trending story

Google DeepMind and UChicago paper finds that AI agents' speaking style affects group policy

1 reports1 story updatesUpdated Nov 18 00:22

Understand this story

AI overview

Google DeepMind and the University of Chicago published a paper in November 2026, finding that AI agents speaking on behalf of individuals gain greater influence over group policy when they are instructed to write longer messages or to cooperate less. The study built multi-agent systems that first interview individuals and then debate with other agents to set an AI safety budget, ran thousands of agent deliberations, and conducted a preregistered study with 746 participants. The results show that agents pull the budget toward the views of the people they represent, and that people consider these agents fairer than submitting opinions to a single AI system, closing 82% of the fairness gap with direct participation. The authors recommend limiting message length and setting shared behavioral rules when building agent-based deliberation.

AI-generated from public reports

Latest development · Report

A paper by GoogleDeepMind and the University of Chicago finds that AI agents speaking on behalf of individuals have a greater influence on group policy the more verbose or the less cooperative they are

Developments

1 confirmed developments

  1. Google DeepMind and UChicago paper finds AI agent speaking style affects group policy

    Source:Rohan Paul

Reports and discussion timeline

Browse reports and discussion by time. Their evidence levels remain distinct.

Nov 18

  1. Rohan Paul

    A paper by GoogleDeepMind and the University of Chicago finds that AI agents speaking on behalf of individuals have a greater influence on group policy the more verbose or the less cooperative they are

    A new paper by GoogleDeepMind and the University of Chicago finds that AI agents speaking on behalf of individuals gained greater influence over group policy when instructed to write longer messages or to cooperate less. The research built multi-agent systems that first interview individuals and then debate with other agents to set AI safety budgets, ran thousands of agent deliberations, and conducted a preregistered study with 746 participants. The agents pulled budgets toward their constituents' views, and people considered these agents fairer than submitting opinions to a single AI system, closing 82% of the fairness gap with direct participation. The authors recommend limiting message length and setting shared behavioral rules when building agent-based deliberation.

Related stories

  1. Related storiesGoogle DeepMind paper shows prompt injection remains unsolved
  2. Related storiesGoogle DeepMind uses AlphaProof Code to formally verify the Signal protocol
  3. Related storiesOpenAI agents attacked online databases and government websites
  4. Related storiesDeepMind and Columbia propose BTTF agent-based chip simulation debugging method