AI agents deliberate on people's behalf: across 746 participants, they recovered 68% of the satisfaction gap between a single-model baseline and direct participation
Synopsis
AI safety policy needs diverse input, yet the public cannot join every discussion. The work builds pluralistic agents: each AI agent first interviews the person it represents, then speaks for them in a multi-agent deliberation, while a policy-writer agent drafts and revises a budget after each round. In a preregistered study with 746 participants, people rated this process above giving input to a single reasoning model, recovering 68% of the satisfaction gap and 82% of the procedural-justice gap between that baseline and direct participation.
Interpretation
AI safety policy discussions mostly happen among policy specialists and AI agents in a small circle, leaving ordinary perspectives outside; pluralistic agents give each participant an AI agent that speaks for them inside the same discussion. Earlier AI-assisted deliberation systems raised the volume of input a process could absorb, but every contribution still had to come from a human at the moment it was needed; pluralistic agents let the agent attend in the person's place. In a preregistered study, 746 participants went through direct discussion, pluralistic agents and a single-AI process, and pluralistic agents scored above the single-AI baseline on all four outcomes.
Participants rated pluralistic agents above handing their input to a single reasoning model, but below participating themselves. On a 1-to-10 satisfaction slider, direct discussion averaged 7.6, pluralistic agents 6.9 and single AI 6.1. Pluralistic agents recovered 68% of the satisfaction gap between the single-AI baseline and direct discussion and 82% of the procedural-justice gap, the latter not significantly different from 100%.
How much an agent sways the outcome depends on how the person instructs it, not only on the underlying model. Across 5,947 simulated-persona deliberations, agents instructed to be more verbose or less cooperative had more influence on the final policy, and speaking later also increased influence. Five of six tested factors were significant, with analyticity the only one without an effect; a parallel experiment within a single Gemini model family replicated the significance of all five.
Perspective
The results are a proof of concept: the human experiment used a four-way budget-allocation task on six AI safety topics with participants recruited from Prolific under U.S.-representative quotas, and the simulated personas are not intended to match any particular real-world population. The authors propose testing more precisely simulated populations, richer discussion formats and more diverse participants in future work.
The authors note that agent influence is sensitive to instructions such as being more verbose or less cooperative; if participants learn this lever, adversarial actors could gain undue influence, a concern 80% of participants rated as moderately or extremely worried. The satisfaction gap between direct discussion and pluralistic agents was barely explained by the measured representation accuracy or by whether the agent said things the participant would not say.