Research map / Large language models / AI and society
AI-Mediated Communication Can Steer Collective Opinion
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.
Shows that open-weight LLMs introduce directional opinion biases when editing human posts, models how such AI-mediated bias propagates through Friedkin-Johnsen opinion dynamics on SNAP networks, and audits Grok's "Explain this post" on X, tracing a pro-life bias to one prompt guideline.
Problem
How biases that AI systems introduce when mediating human-to-human communication (editing or contextualizing posts) affect collective opinion in social networks, and whether platform design choices control these biases.
Approach
Extends the Friedkin-Johnsen opinion dynamics model with an AI transformation f sitting between users, showing analytically that AI bias is amplified through the network (equilibrium shift can exceed one-off bias), validated with LLM edits of human posts and a prompt-level audit of Grok.
Key results
- All four open-weight LLMs introduce statistically significant directional biases on contested topics when editing human posts; e.g., gemma-3-12b-it biases toward 'in favor' on all topics except atheism.
- Under simulated AI adoption on the Twitter SNAP network, the long-run average opinion shift can be up to 9.2 times the average one-off AI bias.
- Grok's 'Explain this post' produces pro-life-leaning claims: support bias 0.24 (95% CI [0.09, 0.38]); the third guideline ('challenging mainstream narratives') accounts for significant increases in both support (0.20) and opposition (0.05) biases.
Surprising findings
- On atheism, LLMs express positive opinions yet introduce biases against atheism when editing posts, so directly expressed opinion benchmarks miss subtle biases.
- Under non-linear AI transformations, individual opinions may not converge in the directed Twitter and Google Plus networks, although the average opinion stabilizes.
Limitations
- Opinion measurement relies on embedding-based classifier ensembles and a single LLM judge (gpt-5.4), not human annotation.
- Theory limited to Friedkin-Johnsen with linear AI transformation; non-linear case only empirical and convergence of individual opinions not guaranteed.
- Grok audit uses only 78 posts from SemEval and one model version, and omits the mediation in real deployments.
Experimental setup
Datasets: UKP Sentential Argument Mining Corpus SemEval-2016 Task 6 SNAP Twitter SNAP Facebook SNAP Google Plus
Benchmarks: UKP Sentential Argument Mining Corpus SemEval-2016 Task 6
Metrics: directional bias (posterior intercept) accuracy macro F1 support bias opposition bias long-run average opinion shift
Baselines: Friedkin-Johnsen model (no AI transformation)
Builds on: Friedkin-Johnsen model Hegselmann-Krause model Nadaraya-Watson kernel regression
Closest papers by meaning
- DEBATE: A Large-Scale Benchmark for Evaluating Opinion Dynamics in Role-Playing LLM Agents (NeurIPS 2026 · similarity 0.87)
- Position: Evolving AI Collectives Enhance Human Diversity and Enable Self-Regulation (ICML 2024 · similarity 0.86)
- Political Neutrality as Balanced Approval: A Large-Scale Human Evaluation of AI Responses (NeurIPS 2026 · similarity 0.86)
- Expressive and Flexible Simulation of Information Spread Strategies in Social Networks Using Planning (AAAI 2024 · similarity 0.86)
- Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond) (NeurIPS 2025 · similarity 0.85)
- Diverse Yet Biased: Towards Mitigating Biases in Generative AI (Student Abstract) (AAAI 2024 · similarity 0.85)
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.