On November 11, 2026, researchers from Google DeepMind and the University of Chicago published a preprint on arXiv titled Scaling public input on AI policy with pluralistic agents. The study introduces and evaluates AI agent representatives that deliberate on behalf of individual citizens to scale public consultation on artificial intelligence governance. For research administrators, ethics boards, and policy researchers, this work provides empirical benchmarks on how autonomous proxies deliberate, while surfacing operational risks regarding prompt sensitivity and participatory agency.
What happened
Public consultation on artificial intelligence regulation often struggles with low civic engagement, complex technical barriers, and the inability of institutions to disclose confidential technical details to broad audiences. To address these bottlenecks, a research team led by Jacy Reese Anthis, Ken Burke, David Madras, Crystal Qian, Meredith Ringel Morris, Logan Cross, Caroline Pantofaru, and Utsav Prabhu designed and evaluated “pluralistic agents”—autonomous software representatives tasked with interviewing individuals and deliberating on their behalf.
The system operates in three stages: first, an agent interviews a person or persona to capture their qualitative values, core priorities, and trade-offs; second, the agent participates in structured multi-agent discussions alongside peer representatives; third, a designated policy-writing agent synthesizes the consensus into concrete rules. By giving each agent one turn per round under anonymized aliases, the process seeks to prevent the status, demographic, and speaking-time imbalances that frequently undermine human deliberations.
Methodology and key findings
The investigators tested the framework across three experimental configurations, combining simulated personas and preregistered human subject evaluations:
- Open-ended swarm governance: In 1,000 multi-agent deliberations involving 368 demographic personas, agents drafted rules for multi-agent swarm oversight across software, laboratory research, customer management, logistics, and public administration. Personas identified a rule matching their most-desired priority in 90% of cases (4,515 of 5,000) for their own group compared to 69% (3,431 of 5,000) for an outside group (a 21.68 percentage point increase, p less than 0.001). Personas assigned an average satisfaction score of 4.65 out of 5 to their own group’s policy versus 3.97 for another group’s policy.
- Policy influence and prompt sensitivity: Across 5,947 deliberations using conjoint analysis, the authors systematically varied six agent inputs: verbosity (5 to 100 words), cooperativeness (1 to 10 scale), analyticity, quantitative preference expression, speaking order, and underlying model architectures (including Claude-Opus-5, GPT-5.6-Sol, GLM-5.2, Kimi-K3, Mistral-Large, and Gemini-3.6-Flash). Less cooperative, highly assertive agents and more verbose agents exerted significantly greater influence over the final policies, whereas conciliatory agents were frequently sidelined.
- Preregistered human subject study: Evaluating 746 verified human participants recruited through Prolific across six AI safety topics, the researchers compared pluralistic agents against direct human-agent discussion and a single-model reasoning baseline. While participants expressed the highest satisfaction with direct personal participation, pluralistic agents recovered 68% of the satisfaction gap and 82% of the procedural justice gap between the single-AI baseline and direct discussion. When asked to choose between pluralistic agent proxies and direct deliberation with other humans, participants split evenly at 50-50.
Why it matters for research practice
As universities, funding agencies, and standards bodies navigate frontier AI safety and governance, traditional consultation mechanisms struggle to keep pace with rapid deployment cycles. Pluralistic agent frameworks demonstrate how institutions can solicit granular public input across high-dimensional technical choices, such as computational safety thresholds and autonomous research laboratory boundaries, without requiring participants to master specialized engineering jargon.
For institutional governance bodies, including AI governance and ethics oversight committees, the study highlights both opportunities and structural cautions. The finding that verbosity and cooperativeness dictate policy outcomes indicates that prompt phrasing functions as de facto procedural rulemaking. Without standardized prompt calibration and transparent moderation algorithms, agent-driven consultations risk skewing outcomes toward more assertive delegates.
Furthermore, university Research Ethics Boards (REBs) reviewing participatory research protocols must recognize the ethical nuances of algorithmic proxy representation. While agents expanded democratic throughput, human participants reported concerns regarding personal agency, noting that delegating deliberation prevented them from engaging in direct social learning and updating their own viewpoints during debate.
Governance risks and institutional implications
The authors emphasize that pluralistic agents should function as amplifiers of public engagement rather than autonomous replacements for democratic institutions. While direct human participation achieved an average attitude shift of 15.4 percentage points in budget allocations before versus after debate, pluralistic agent deliberation produced a comparable 13.9 percentage point shift, indicating meaningful policy responsiveness.
Institutional researchers adopting multi-agent deliberative tools should implement safeguards against prompt manipulation, establish clear audit trails for agent interviews, and ensure that human stakeholders retain ultimate ratification authority over algorithmic consensus policies.
Sources
- Anthis, J. R., Burke, K., Madras, D., Qian, C., Morris, M. R., Cross, L., Pantofaru, C., and Prabhu, U. (2026). Scaling public input on AI policy with pluralistic agents. arXiv preprint arXiv:2611.16028. https://arxiv.org/abs/2611.16028
