Research map / Large language models / AI and society
Should You Use LLMs to Simulate Opinions? Quality Checks for Early-Stage Deliberation
Terrence Neumann, Maria De-Arteaga, Sina Fazelpour
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| fine-tune on 40 human-labeled claims instead of backstory prompting (ablation) | GPT-3.5c opinion simulation · TopicMisinfo QC2a/QC2b | share of topics passing: QC2a 50% -> 100%; QC2b 70% -> 100% | Improved |
| fine-tune on small human data instead of backstory prompting (ablation) | GPT-3.5c opinion simulation · TopicMisinfo QC1a weak consistency | share of thresholds passed: 44% -> 19% | Hurt |
| use nearest-neighbor in-context examples (ablation) | GPT-3.5c opinion simulation · TopicMisinfo QC1a | share of thresholds passed: 44% -> 88% | Improved |
| use nearest-neighbor in-context examples (ablation) | GPT-3.5c opinion simulation · TopicMisinfo QC1b strong test | share of topics passed: 25%/50% -> 0% | Hurt |
| use random-sample in-context examples (ablation) | GPT-3.5c opinion simulation · TopicMisinfo QC1a/QC1b | share passed: QC1a 44% -> 56%; QC1b 25% -> 50% vs Prompt 1 | Mixed |
| fine-tuning on small-sample data for strong consistency (ablation) | GPT-3.5c opinion simulation · TopicMisinfo QC1b | share of topics passed: 25% -> 25% | No clear effect |
| move from GPT-3.5 to GPT-4/GPT-4.1/GPT-5-mini/LLaMA3/Mistral/Titan | backstory prompting · TopicMisinfo QC2b divisive topics | weighted gender-gap deviation: Abortion GPT-3.5a 1.65 -> GPT-4 -1.70 | Hurt |
| use GPT-3.5 family models for non-divisive topics | backstory prompting · TopicMisinfo Gold QC2a | simulated gender gap: GPT-3.5c 0.72*** vs GPT-4 0.00 | Hurt |
| switch to reasoning model GPT-5-mini | backstory prompting · TopicMisinfo QC1 weak/strong | consistency tests passed | No clear effect |
| vary neutral prompt wording across four base prompts (ablation) | backstory prompting · TopicMisinfo QC1 weak/strong | consistency tests passed: 13/64 pass p0=0.8; 2/64 pass p0=0.9 | Mixed |
Closest papers by meaning
- Chatbot Arena: An Open Platform for Evaluating LLMs by Human Preference (ICML 2024 · similarity 0.90)
- Can LLMs Express Their Uncertainty? An Empirical Evaluation of Confidence Elicitation in LLMs (ICLR 2024 · similarity 0.90)
- Uncertainty Quantification for LLM-Based Survey Simulations (ICML 2025 · similarity 0.89)
- Bridging Human and LLM Judgments: Understanding and Narrowing the Gap (NeurIPS 2025 · similarity 0.88)
- Neither Valid nor Reliable? Investigating the Use of LLMs as Judges (NeurIPS 2025 · similarity 0.88)
- More Is Not More: What Matters for Diversity in LLM Opinions? (NeurIPS 2026 · similarity 0.88)
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.
