Research map / Large language models / AI and society
AI Debate Aids Assessment of Controversial Claims
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Compares AI debate against single-advisor consultancy for human and LLM judges with mainstream or skeptical COVID-19 and climate change beliefs on controversial factuality claims.
Problem
Scalable oversight relies on judges who may have strong prior beliefs; it is unclear whether debate can guide biased judges toward truth on contested factuality claims better than consultancy, and whether persona-based LLM judges can supervise as well as humans.
Approach
Two GPT-4o debaters argue opposing sides with reference sources while human judges with mainstream or skeptical beliefs evaluate claims; the study compares this to a single-advisor consultancy and tests persona-conditioned LLM judges against human judges and default LLM judges.
Key results
- Debate improved human judge final accuracy to 70.1% vs 60.0% for consultancy on COVID-19 claims (about 10% overall advantage, p<0.01); +15.2% for mainstream-belief judges.
- Persona-based GPT-4o judges with debate reached 78.5% accuracy (public personalized), above human judges at 70.1%.
- Debate yielded better calibration (Brier 0.231 vs 0.335 for consultancy) and fewer harmful belief reversals for mainstream judges (8.6% vs 22.9%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use two opposing AI advisors instead of a single advisor | Human judges evaluating COVID-19 claims · COVID-19 claims | final accuracy: 60.0% -> 70.1% | Helped |
| Use two opposing AI advisors instead of a single advisor | Human judges evaluating climate change claims · Climate change claims | accuracy: 68.3% -> 72.1% | Helped |
| Use two opposing AI advisors instead of a single advisor | Mainstream human judges evaluating COVID-19 claims · COVID-19 claims, mainstream judges | final accuracy: 55.9% -> 71.1%; 15.2% | Helped |
| Use two opposing AI advisors instead of a single advisor | Skeptical human judges evaluating COVID-19 claims · COVID-19 claims, skeptical judges | final accuracy: 64.3% vs 69.0%; 4.7% | Helped |
| Use debate rather than consultancy | Human judges evaluating COVID-19 claims · COVID-19 claims | Brier score: 0.335 vs. 0.231 | Helped |
| Use debate rather than consultancy | Mainstream human judges evaluating COVID-19 claims · COVID-19 claims, mainstream judges | harmful belief reversals: 22.9% vs 8.6% | Helped |
| Use debate rather than consultancy | Skeptical human judges evaluating COVID-19 claims · COVID-19 claims, skeptical judges | truth-aligned confidence shift: +6.9 -> +14.6 | Helped |
| Use debate rather than consultancy | Mainstream human judges evaluating COVID-19 claims · COVID-19 claims, mainstream judges | truth-aligned confidence shift: -12.6 -> +15.3 | Helped |
| Add a human-like persona to the LLM judge, with debate agents aware of it | GPT-4o judge · COVID-19 claims | accuracy: 69.8% -> 78.5% | Helped |
| Add a human-like persona to the LLM judge, with debate agents unaware of it | GPT-4o judge · COVID-19 claims | accuracy: 69.8% -> 77.3% | Helped |
| Use a persona-based LLM judge instead of a human judge | Debate on COVID-19 claims · COVID-19 claims | accuracy: 70.1% -> 78.5% | Helped |
| Personalize the judge persona in consultancy | LLM judges modeling human decisions · COVID-19 claims | human-LLM agreement: 52.2% -> 64.7% | Helped |
Surprising findings
- Consultant agreement with judges' initial beliefs did not reliably improve accuracy, and consultancy decreased truth alignment for mainstream judges.
- Default LLM judges without personas performed comparably to human judges with debate (GPT-4o 69.8%).
Limitations
- Controlled simulation of debate and consultancy; in-the-wild deployment may differ.
- Annotators are U.S.-based English speakers, not census-representative.
- Skeptical-group debate advantage over consultancy was not statistically significant (p=0.2).
Experimental setup
Datasets: CheckCOVID CORD-19 Climate-Fever
Benchmarks: CheckCOVID Climate-Fever
Metrics: accuracy calibration (Brier score) belief change human-LLM agreement
Baselines: consultancy non-personalized LLM judges
Builds on: AI safety via debate consultancy
Closest papers by meaning
- Collaborative Disagreement Resolution for Scalable Oversight (ICML 2026 · similarity 0.86)
- On scalable oversight with weak LLMs judging strong LLMs (NeurIPS 2024 · similarity 0.85)
- Beyond Detection: Exploring Evidence-based Multi-Agent Debate for Misinformation Intervention and Persuasion (AAAI 2026 · similarity 0.83)
- Can AI Agents Synthesize Scientific Conclusions? (NeurIPS 2026 · similarity 0.83)
- Debating with More Persuasive LLMs Leads to More Truthful Answers (ICML 2024 · similarity 0.83)
- When Debate Helps: Proposal Supply and Verification-Aware Readout in Multi-Agent Reasoning (NeurIPS 2026 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.