💬 NLP

Can AI Truly Represent Your Voice in Deliberations? A Comprehensive Study of Large-Scale Opinion Aggregation with LLMs

This paper introduces DeliberationBank, a large-scale human-annotated dataset, and DeliberationJudge, a fine-tuned model that outperforms LLMs in evaluating deliberation summaries, to systematically identify and address persistent biases such as the underrepresentation of minority perspectives in AI-driven public deliberations.

Original authors: Shenzhe Zhu, Shu Yang, Michiel A. Bakker, Alex Pentland, Jiaxin Pei

Published 2026-03-23

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Shenzhe Zhu, Shu Yang, Michiel A. Bakker, Alex Pentland, Jiaxin Pei

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a massive town hall meeting where 3,000 people are shouting their opinions about important issues like taxes, AI, or healthcare. The goal is to take all that noise and turn it into a single, clear, and fair summary that a politician can actually use to make decisions.

This paper asks a big question: Can Artificial Intelligence (AI) do this job without leaving anyone out?

Here is the story of their research, broken down into simple parts with some creative analogies.

1. The Problem: The "Echo Chamber" AI

In the past, if you wanted to summarize 3,000 opinions, you'd need a team of human editors. It's slow and expensive. So, people started using AI (Large Language Models, or LLMs) to do it.

But there's a catch. Think of AI like a popular kid in high school. When the popular kid summarizes what the whole class thinks, they tend to repeat what the majority said because that's what they hear the most. They often miss the quiet kid in the back or the person with a weird, unique idea.

The researchers found that current AI summarizers are great at sounding smart, but they often ignore minority viewpoints. If 90% of people want "Option A" and 10% want "Option B," the AI summary might say, "Everyone wants Option A," completely erasing the 10%. In a democracy, that's a big problem.

2. The Solution: Building a "Fairness Test" (DeliberationBank)

To fix this, the researchers (from Stanford, MIT, and others) built a giant testing ground called DeliberationBank.

The Setup: They gathered 3,000 real human opinions on 10 different hot topics.
The Test: They asked 18 different AI models to summarize these opinions.
The Judges: They hired 4,500 real humans to grade those AI summaries. The humans didn't just ask, "Is this a good summary?" They asked four specific questions:
1. Representativeness: Did this summary capture my specific opinion?
2. Informativeness: Did it give me new, useful details?
3. Neutrality: Did it sound fair, or was it biased?
4. Policy Approval: Would I trust a politician to use this to make a law?

3. The "Super Judge" (DeliberationJudge)

Here is the tricky part: You can't hire 4,500 humans every time you want to test a new AI. That's too expensive. So, the researchers tried using other AIs to act as judges.

The Bad News: When they asked big, fancy AI models to grade the summaries, the AI judges were inconsistent. They were like unreliable referees who changed the rules every time they blew the whistle. They didn't agree with the real humans very well.

The Good News: The researchers trained a special, smaller AI model called DeliberationJudge.

Think of this model as a veteran sports coach who has watched thousands of games. It was trained specifically on the feedback from the 4,500 real humans.
Result: This coach is incredibly accurate. It agrees with human opinions much better than the big, general-purpose AIs. Plus, it's super fast and cheap to run (100 times faster than the big models).

4. What They Discovered

Using their new "Super Judge," they tested 18 different AI models and found some surprising things:

Bigger isn't always better: The most expensive, massive AI models weren't necessarily the fairest. Some smaller, specialized models did a better job of staying neutral.
The "Minority Blindspot": Almost every AI model failed to represent minority opinions well. Whether the minority opinion was a "self-reported" feeling of being different, or an "objectively rare" idea, the AI tended to smooth it over and make it disappear.
Two types of "Minorities": They found that some people feel like they are in the minority (even if their words are common), while others have actually unique words (even if they don't realize it). The AI missed both types.

5. The Takeaway

This paper is like a quality control report for the future of democracy.

It tells us that while AI is a powerful tool for summarizing public opinion, we can't just trust it blindly. If we use AI to summarize what the public thinks, we risk creating a "fake reality" where the majority voice drowns out everyone else.

However, the researchers also gave us the tools to fix it:

DeliberationBank: A massive dataset to keep testing AI fairness.
DeliberationJudge: A specialized tool to ensure AI summaries are actually fair and representative before we let them influence real-world policy.

In short: AI can help us listen to the crowd, but only if we give it a "fairness filter" to make sure it doesn't just hear the loudest voices.

1. Problem Statement

Large-scale public deliberations (e.g., citizens' assemblies, parliamentary debates) generate thousands of free-form opinions that must be synthesized into neutral, representative summaries for policy-making. While Large Language Models (LLMs) offer a scalable solution for this summarization, two critical challenges remain:

Representation Bias: LLMs tend to over-represent majority viewpoints while systematically under-representing minority perspectives, raising fairness concerns in high-stakes democratic contexts.
Evaluation Reliability: Current practices rely on "LLM-as-a-judge" paradigms to evaluate summaries. However, general-purpose LLMs show weak alignment with human judgments, exhibit instability, and are prone to their own biases, making them unreliable for benchmarking deliberation tools.

The paper addresses the central question: Can LLMs produce summaries that are truly representative, informative, and neutral for policy use, and how can we evaluate them reliably at scale?

2. Methodology

The authors propose a comprehensive framework consisting of a new dataset, a specialized evaluation model, and a rigorous benchmarking study.

A. DELIBERATIONBANK (Dataset)

A large-scale, human-grounded benchmark created through a three-stage pipeline involving 7,500 unique U.S. participants:

Stage 1 (Opinion Collection): 3,000 free-form opinions were collected across 10 deliberation questions (5 binary policy questions, 5 open-ended societal questions) from 300 representative participants per question.
Stage 2 (Summary Generation): 750 summaries were generated using 5 different LLMs (including GPT-5, Claude-4, Gemini-2.5, DeepSeek, and Qwen3) across varying input subset sizes (10 to 300 opinions) to test scalability.
Stage 3 (Human Annotation): 4,500 human annotations were collected from a distinct pool of 450 participants per question. Each annotator wrote their own opinion and evaluated summaries on four dimensions:
1. Representativeness: Semantic capture of the individual's opinion.
2. Informativeness: Diversity and detail of information.
3. Neutrality: Absence of bias language.
4. Policy Approval: Suitability for decision-making.
  Annotations included both absolute rating (1–5 Likert scale) and pairwise comparison tasks.

B. DELIBERATIONJUDGE (Evaluator)

To overcome the limitations of general-purpose LLM judges, the authors developed DELIBERATIONJUDGE:

Architecture: A fine-tuned DeBERTa-v3-base model.
Training Strategy: Supervised Fine-Tuning (SFT) on the 4,500 human annotations. It unifies rating and comparison tasks by normalizing labels into a continuous $[0, 1]$ space.
Objective: To predict a 4-dimensional score vector that aligns closely with human preferences while maintaining high inference efficiency.

C. Benchmarking Protocol

The authors evaluated 18 LLMs (ranging from 0.6B to 671B parameters, including proprietary and open-weight models) using the DELIBERATIONBANK dataset. They analyzed performance across:

Model size and architecture (Dense vs. MoE).
Input scale (number of opinions summarized).
Question type (Binary vs. Open-ended).
Minority representation (Subjective self-reports vs. Objective semantic outlier detection).

3. Key Contributions

DELIBERATIONBANK: The first large-scale, human-grounded dataset for deliberation summarization, containing 3,000 opinions and 4,500 multi-dimensional evaluations.
DELIBERATIONJUDGE: A specialized, lightweight evaluator that outperforms general-purpose LLMs in both alignment with human judgment and computational efficiency.
Systematic Bias Analysis: A rigorous identification of systematic under-representation of minority viewpoints in LLM-generated summaries, distinguishing between subjective (self-perceived) and objective (semantically unique) minorities.

4. Key Results

A. Evaluation Performance

Human Alignment: General-purpose LLMs (even large ones like GPT-5) show weak correlation with human judgments (Pearson/Spearman $\rho < 0.4$ ).
DELIBERATIONJUDGE Superiority: The fine-tuned DeBERTa model achieves a Spearman correlation of 0.70 with human judgments (vs. ~0.20 for the best LLM judge) and is 100x more efficient (0.03s vs. >8s per item).
Rank Correlation: DELIBERATIONJUDGE shows near-perfect rank alignment on Informativeness ( $\rho=1.00$ ) and high alignment on Neutrality and Policy Approval ( $\rho=0.90$ ).

B. Summarization Benchmarks

Model Performance: Top-tier models (GPT-5, Claude-4) achieved average scores of ~0.62 across dimensions. Performance gaps between models were modest (<0.1), suggesting most models capture the "gist" similarly but struggle with nuance.
Scalability Limits: Summarization quality improves as input size increases from 10 to ~100 opinions but plateaus thereafter. Models fail to effectively leverage larger input sets, indicating a need for better aggregation mechanisms.
Topic Sensitivity: Models perform better on binary questions (constrained stances) than open-ended questions (diverse views).

C. Minority Representation Bias (Critical Finding)

Systematic Under-representation: Across all 18 models, summaries consistently received lower representativeness scores for minority opinions compared to majority opinions.
- Subjective Minority: Opinions where users self-identified as holding a minority view.
- Objective Minority: Opinions identified as semantic outliers via Local Outlier Factor (LOF) analysis.
The Gap: The bias gap was significant (up to 0.08 in binary topics). Crucially, the overlap between subjective and objective minorities was <10%, indicating they represent distinct phenomena:
- Subjective minorities often feel isolated due to personal context or uncertainty, even if their semantic content is common.
- Objective minorities contain rare semantic arguments or unique reasoning patterns that models deprioritize.
Random Control: Randomly sampled groups showed negligible bias, confirming that the under-representation is specific to actual minority perspectives, not an artifact of sample size.

5. Significance and Implications

Reliability in Democratic AI: The paper demonstrates that current "LLM-as-a-judge" approaches are insufficient for evaluating tools used in democratic processes due to poor alignment with human values and inherent biases.
Fairness in Summarization: It provides empirical evidence that LLMs systematically silence minority voices, a critical failure mode for AI in policy-making.
Scalable Evaluation Framework: By introducing DELIBERATIONJUDGE, the authors provide a cost-effective, scalable method to audit and improve deliberation systems without relying on expensive human evaluation for every iteration.
Future Directions: The work suggests that future deliberation AI must incorporate explicit fairness constraints, diversity-aware objectives, and specialized evaluation metrics to ensure equitable representation of all stakeholders.

In conclusion, while LLMs show promise for summarizing public deliberation, they currently lack the necessary fairness and evaluation reliability for high-stakes deployment. The proposed DELIBERATIONBANK and DELIBERATIONJUDGE offer the foundational tools required to build more representative, equitable, and policy-relevant AI systems.