Agora: Teaching the Skill of Consensus-Finding with AI Personas Grounded in Human Voice

The paper introduces Agora, an AI-powered platform that leverages LLMs to simulate diverse human perspectives on policy issues, enabling users to practice consensus-building and demonstrating through a preliminary study that access to authentic voice explanations significantly enhances problem-solving skills and the quality of collective decisions compared to viewing aggregate data alone.

The Gist

Imagine you are trying to bake the perfect cake for a huge party. You know everyone has different tastes: some love chocolate, some hate nuts, and some are on a strict diet. If you just look at a list of "50% like chocolate, 50% hate nuts," you might guess the recipe, but you won't really understand why.

Agora is a new digital tool designed to teach people how to find that "perfect cake" (or in the real world, the perfect public policy) by turning the boring list of opinions into a lively, interactive conversation.

Here is a simple breakdown of how it works, why the researchers built it, and what they found.

The Problem: We Are Bad at Listening

The researchers believe that being a good citizen isn't something you are born with; it's a skill you have to practice, just like playing the guitar or solving math problems.

However, most of us don't get to practice this skill. Real-life town halls or government meetings are rare, and when we do see opinions online, we usually just see a "poll result" (e.g., "60% say Yes"). We don't hear the story behind the "Yes." We don't know why a neighbor supports a new law. Without understanding the "why," it's hard to find common ground.

The Solution: A "Voice-First" Video Game for Policy

The team built Agora, an AI-powered platform that acts like a simulation game for democracy.

1. The Cast of Characters (The Personas):
   Instead of using fake computer bots, the researchers first interviewed 90 real people from across the US. They asked them about their lives and their views on tough topics like "How much should the minimum wage be?" or "Should companies hire locals over foreigners?"

   The AI then turned these real interviews into digital avatars. These aren't just static pictures; they are "personas" that can speak in the real voices of the people interviewed.

2. The Game:
   You log in and are asked to write a policy (like a new law).

   • The Visuals: On your screen, you see a line of avatars. As you type your policy, the avatars move left or right based on how much they support your idea.
   • The Magic: If you click on an avatar, you don't just see a text summary. You hear a 60-second audio clip of that real person explaining why they support or oppose your idea. They might say, "I support this because I lost my job when my factory moved," or "I oppose this because I need to hire my cousin who is an immigrant."

3. The Feedback Loop:
   You tweak your policy, hit "Calculate," and watch the avatars move again. You might realize, "Oh! If I add a small exception for small businesses, suddenly 20 more people support me!" You can then click those new supporters to hear their stories.

It's like having a personal debate coach that instantly shows you how changing one sentence in your proposal changes the minds of real people, and gives you the emotional context to understand them.

The Experiment: Does It Work?

The researchers tested this with 44 university students. They split them into two groups:

• Group A (The Full Experience): Could see the avatars, hear the real voices, and read the stories.
• Group B (The Control): Could see the same avatars moving left and right, but they were just generic icons with no voices or stories. They only saw the numbers.

The Results:

• Group A felt like they were learning more. They reported feeling more confident in their problem-solving skills.
• Group A spent more time thinking deeply about the issue before talking to others (a skill called "internal deliberation").
• The Output: When asked to write a final consensus statement, Group A wrote much better, clearer, and more specific policies. Group B tended to write vague, generic statements just to get a high score, without really thinking through the details.

The Big Takeaway

The study suggests that knowing why people disagree is just as important as knowing that they disagree.

When you hear a real human voice explaining their struggle or hope, it's harder to dismiss them as just a "statistic." Agora teaches people that finding consensus isn't about winning an argument; it's about listening to the stories behind the opinions and adjusting your ideas to fit the reality of other people's lives.

Why This Matters

We can't invite every citizen in the world to a town hall meeting. But if we can use AI to scale up this kind of "listening practice," we might be able to teach a whole generation how to navigate disagreement, weigh different values, and build a better future together.

In short: Agora turns the dry math of voting into a human story, helping us learn that the best solutions come from understanding the people behind the numbers.

Technical Summary

Here is a detailed technical summary of the paper "Agora: Teaching the Skill of Consensus-Finding with AI Personas Grounded in Human Voice."

1. Problem Statement

Deliberative democratic theory posits that civic competence—the ability to navigate disagreement, weigh competing values, and reach collective decisions—is a learned skill rather than an innate trait. However, traditional methods for cultivating these skills (e.g., citizens' assemblies, deliberative polls) are limited in scale, reaching only a small fraction of the population.

The core challenge addressed by this research is: How can technology scale civic education to allow individuals to practice deliberative skills at a population level? Existing tools often present aggregate data or static viewpoints, failing to provide the iterative feedback loop necessary for skill acquisition. There is a need for a platform that transforms passive exposure to opinions into active skill practice by allowing users to propose policies, hear authentic diverse perspectives, and receive immediate feedback on how their proposals affect support.

2. Methodology

A. System Architecture: Agora

The authors developed Agora, an AI-powered platform designed to simulate a deliberative environment using Large Language Models (LLMs) and voice-based data.

• Data Foundation: The system is grounded in a corpus of 90 semi-structured voice interviews conducted via Prolific with US residents (balanced for political ideology). The interviews covered personal experiences and beliefs regarding policy issues (minimum wage, domestic vs. foreign hiring).
• AI Interviewer: An automated voice-to-voice system (using Whisper for STT, GPT-4o for generation, and TTS for synthesis) conducted the initial interviews, prioritizing low latency and natural conversation flow.
• Backend Processing Pipeline:
  1. Support Prediction: GPT-4.1 analyzes interview transcripts against a user's drafted policy to predict the interviewee's level of support (0–100%). The model is prompted to provide reasoning and confidence scores before outputting the prediction to enhance accuracy.
  2. Audio Medley Generation: For each interviewee and policy iteration, GPT-4.1 identifies specific transcript segments containing personal experiences and reasoning that support the predicted stance. These are synthesized into 60–90 second audio medleys.
  3. Dynamic Feedback Loop: When a user revises a policy and clicks "Calculate," the system re-runs the entire pipeline: updating support predictions, regenerating relevant audio medleys, and shifting the visual position of avatars along a support spectrum.

B. User Interface

• Visualization: Users see a horizontal axis populated by avatars (generated via GPT-5 based on demographic data) representing the interviewees. Their position indicates predicted support for the current policy.
• Interaction: Clicking an avatar plays the personalized audio medley and displays a text summary, explaining why that individual supports or opposes the policy.
• Goal: Users aim to iteratively revise their policy text to maximize the aggregate support score across the population.

C. Experimental Design

• Participants: 44 university students in the US.
• Task: Draft optimal policies for two topics: (1) Minimum wage levels, and (2) Domestic vs. foreign hiring priorities.
• Conditions:
  • Treatment Group: Used the full Agora interface with access to voice explanations, avatar interactions, and dynamic feedback.
  • Control Group: Used the same drafting tool but saw avatars as generic, non-interactive icons. They could see the aggregate distribution of support change but not the reasons behind the shifts.
• Evaluation Metrics:
  • Self-Reported: Problem-solving skills, interest, internal deliberation (using scales from Shroff et al. and Weinmann).
  • Output Quality: Consensus statements were evaluated using an LLM-as-a-judge approach, scoring for coherence, clarity, evidence integration, specificity, and balanced treatment of uncertainty.

3. Key Contributions

1. A Novel Platform for Civic Skill Practice: Agora introduces a scalable interface that moves beyond static opinion visualization to dynamic, voice-grounded deliberation. It operationalizes the concept of "learning by doing" in a digital civic space.
2. Grounding AI in Human Voice: Unlike many AI deliberation tools that rely on synthetic text or abstract summaries, Agora utilizes authentic voice clips and personal narratives. This preserves tone and narrative context, which the authors argue is crucial for humanizing polarized contexts and bridging moral divides.
3. Dynamic Feedback Mechanism: The system demonstrates a real-time feedback loop where policy changes immediately alter the "opinion landscape" and the reasoning provided by personas, mimicking the iterative nature of real-world negotiation.
4. Empirical Validation of "Why" over "What": The study provides early evidence that understanding the reasons behind opposing views (via voice/personas) is more effective for skill development than merely seeing the distribution of views.

4. Results

• Self-Reported Learning: Participants in the Treatment condition reported significantly higher levels of:
  • Problem-solving skills.
  • Topic interest.
  • Internal deliberation ("deliberating within").
  • Perceived relevance and timeliness of feedback.
• Consensus Statement Quality: The consensus statements generated by the Treatment group were rated higher by the LLM-judge across all dimensions, particularly in specificity and balanced treatment of uncertainty.
  • Control Group Behavior: Tended to produce vague, generic statements designed to garner broad support but lacking actionable detail.
  • Treatment Group Behavior: Produced clearer, more coherent, and specific policies, likely due to engaging with the specific constraints and reasoning of the personas.
• Prediction Accuracy: The LLM's support prediction model achieved an average accuracy of 82% when validated against the interviewees' actual pre-survey votes.

5. Significance and Future Directions

Significance:
The study suggests that mediated engagement with authentic perspectives can effectively foster deliberative skills at a scale impossible for traditional face-to-face assemblies. By shifting the focus from "what people think" to "why they think it," tools like Agora can help citizens develop the perspective-taking capacities required for democratic competence. This offers a potential solution to the scalability crisis in civic education.

Limitations:

• Sample Bias: The study relied on university students, which may not generalize to the broader public or more technical/emotionally charged issues.
• Evaluation Reliance: The use of LLM-as-a-judge for quality assessment has known limitations regarding bias and alignment with human judgment (though a human validation study is planned).
• Simplification: Reducing complex policy preferences to a single support axis oversimplifies the multidimensional nature of real-world trade-offs.
• Ethical Concerns: Issues regarding voice anonymization, user agency over data disclosure, and the potential for LLM biases to shape the organization of perspectives remain unaddressed.

Future Work:
The authors plan to conduct rigorous validation studies with human annotators, expand the participant pool to diverse demographics, and refine the tool to better represent multidimensional policy trade-offs. They also aim to investigate the "active ingredients" of the interface (e.g., voice vs. text) to optimize skill acquisition.