AI proxies can widen policy consultation, but assertive agents may steer the result

Researchers from Google DeepMind and the University of Chicago have tested AI agents that interview people, deliberate as their representatives and draft policy recommendations. The results suggest the approach could broaden public input on AI policy, while leaving a pointed question about whose voice carries most weight.

Google DeepMind Watch analysis

What happened

The study, titled “Scaling public input on AI policy with pluralistic agents”, was published as an arXiv preprint on 11 November, according to CASRAI. The researchers tested the agents in simulated deliberations and a study involving 746 human participants.

Key findings

• Agents often reflected their own group’s priorities
  In 1,000 simulated deliberations, personas’ preferred priority appeared in the resulting rules 90 per cent of the time for their own group, compared with 69 per cent for an outside group.
• How an agent speaks affected its influence
  Across 5,947 deliberations, more verbose and less cooperative agents had greater influence over the final policies; more conciliatory agents were often sidelined.
• People valued direct participation most
  Participants reported the highest satisfaction when taking part themselves. The agent approach recovered 68 per cent of the satisfaction gap and 82 per cent of the procedural-justice gap between a single-model baseline and direct discussion.
• Preference for proxies was evenly split
  When asked to choose between agent representatives and direct deliberation with other people, participants split 50–50.

Why it matters

AI proxies could help institutions collect views on technical policy questions from people who might otherwise find the process difficult to access. But a representative agent does not simply relay a person’s preferences: the study found that traits such as verbosity and cooperativeness affected which views shaped the outcome. That makes the design of prompts and deliberation rules part of the policy process, not a bit of invisible plumbing.

Our read

This is a promising way to widen consultation, not a democratic shortcut with a chatbot-shaped stamp of approval. The results give institutions a reason to test agent-mediated input, and a reason to publish how those agents are instructed and how their conclusions are checked. Human participation and final authority should remain in the room.

What to watch

• Whether further studies reproduce the results with real-world policy consultations.
• How prompt design and speaking rules affect whose preferences reach a final proposal.
• Whether institutions give participants a clear way to review and challenge an agent’s account of their views.

Discussion spark: Should institutions use AI proxies to widen public input on technical policy, or does the risk that assertive agents shape the result outweigh that gain?

not affiliated with, endorsed by, or operated by Google or Google DeepMind