SI Agent News
DevelopingGeminiAgentsPolicySecurity

AI agents could help scale public input on government policy

Researchers from Google DeepMind and the University of Chicago developed a system where AI proxies interview citizens to deliberate on complex regulations.

Ink and watercolour: picture for the story 'AI agents could help scale public input on government policy'
AI-made ink and watercolour for this story. No real people or products are shown.
90%Cases where agents matched a persona's most-desired priority

The short version

Why it matters

This framework allows institutions to gather granular public input on technical topics without requiring every citizen to become a subject matter expert. However, it also reveals that how these agents are prompted can significantly bias the final rules produced.

Scaling public consultation

Public consultation on artificial intelligence regulation often faces significant hurdles. Many people find the technical details of AI governance too complex to navigate, and institutions often struggle to disclose confidential information to the general public. These barriers can lead to low civic engagement and a lack of diverse perspectives in the rulemaking process.

To address these issues, a research team led by Jacy Reese Anthis and colleagues from Google DeepMind and the University of Chicago developed "pluralistic agents." These are autonomous software representatives designed to act on behalf of individual citizens. The goal is to scale public input by allowing software to handle the heavy lifting of deliberation while still representing human values.

As CASRAI reported, the researchers wanted to see if these agents could effectively capture and negotiate the priorities of a broad population. By using AI as a proxy, the system aims to bypass the limitations of traditional human-led consultations. This approach could potentially allow for more inclusive policy-making across a wider range of technical and social issues.

How the agent framework works

The system operates through a structured three-stage process. First, an agent interviews a person or a specific persona to identify their qualitative values and core priorities. This stage is meant to capture the nuances of what an individual cares about most, such as specific trade-offs they are willing to make in a policy environment.

In the second stage, these agents participate in structured multi-agent discussions alongside other peer representatives. To ensure fairness, the system gives each agent only one turn per round and uses anonymized aliases. This design choice is intended to prevent status, demographic, or speaking-time imbalances from undermining the deliberation, which often happens in human-led forums.

The final stage involves a designated policy-writing agent. This specific agent takes the consensus reached during the multi-agent discussions and synthesizes it into concrete rules. By moving from individual interviews to group deliberation and finally to rule synthesis, the framework creates a pipeline for turning personal values into actionable policy.

Testing the agent's influence

The researchers tested the framework using 1,000 multi-agent deliberations involving 368 demographic personas. They looked at various topics, including software oversight, laboratory research, and public administration. The results showed that the agents were quite successful at identifying the most-desired priorities for their specific groups in 90% of the cases.

However, the study also revealed significant risks regarding how these agents behave. The researchers varied inputs like verbosity, cooperativeness, and model architecture to see how they affected the final outcomes. They found that agents that were less cooperative, more assertive, and more verbose exerted significantly greater influence over the final policies than their more conciliatory peers.

This suggests that the way a prompt is written functions as a form of procedural rulemaking. If the prompts for these agents are not carefully calibrated, the final policies might be skewed toward the most assertive delegates. The study highlights that the technical configuration of the AI can determine which voices are heard most loudly in a simulated democratic process.

Human feedback and agency

To validate the system, the team conducted a preregistered study with 746 verified human participants. They compared the pluralistic agents against both direct human-agent discussions and a single-model reasoning baseline. While participants reported the highest satisfaction with direct personal participation, the pluralistic agents recovered a large portion of the satisfaction and procedural justice gap.

When asked to choose between using an AI proxy or engaging in direct deliberation with other humans, participants split evenly at 50-50. This suggests that many people may be open to using AI as a representative in complex policy discussions. It indicates a potential middle ground between traditional town halls and purely automated governance systems.

Despite the high satisfaction scores, some human participants expressed concerns about personal agency. They noted that delegating the deliberation to an agent prevented them from engaging in direct social learning. They felt that by using a proxy, they missed out on the opportunity to update their own viewpoints through direct debate with other people.

Implications for governance

The researchers emphasize that pluralistic agents should be viewed as amplifiers of public engagement rather than replacements for human institutions. The study showed that the agents produced a 13.9 percentage point shift in policy attitudes, which was comparable to the 15.4 percentage point shift seen in direct human debates. This suggests that the framework is capable of producing meaningful policy responsiveness.

For institutions like universities and funding agencies, this framework provides a way to solicit granular input on high-dimensional technical choices. It allows them to gather data on things like computational safety thresholds without requiring every participant to master specialized engineering jargon. This could make it easier for diverse groups to contribute to frontier AI safety and governance.

Moving forward, the authors suggest that institutions must implement safeguards against prompt manipulation. They recommend establishing clear audit trails for how agents are interviewed and ensuring that human stakeholders retain the final authority. Without these protections, the risk remains that algorithmic consensus could be steered by the underlying model architectures or the specific phrasing of the instructions.

What to watch

Try it

Look for public AI policy consultation websites to see if they offer any digital tools for submitting feedback or participating in surveys.

Sources and links

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

Listen to the AI Wire headlines

Listen to the AI Wire

Listen to our daily AI shows every morning: on your commute, in the shower, wherever you are

Listen to the daily shows

First reported by CASRAI · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

More AI news