The Invisible Editor: When AI Sits in the Middle of Every Conversation

Paper: *AI-Mediated Communication Can Steer Collective Opinion* — Stratis Tsirtsis, Kai Rawal, Chris Russell (University of Oxford, Hasso Plattner Institute)

Key points

• The blind spot: Prior work studied LLMs talking *to* people. This paper studies LLMs sitting *between* people—drafting, polishing, or explaining posts (e.g., LinkedIn's "improve your post," X's "Explain this post"). This mediation step is where bias sneaks in.
• Empirical finding: Across 4 open-source LLM families, 13 controversial topics (abortion, gun control, atheism, death penalty, etc.), and two tasks (drafting and improving posts), every tested model introduced directional bias—e.g., outputs leaned toward supporting gun control and opposing atheism. The bias persisted even under explicit instructions to preserve the original viewpoint.
• Theoretical finding: Extending the Friedkin-Johnsen opinion dynamics model with an AI transformation function, the authors prove that at equilibrium:

opinion shift = (AI single-pass bias) × (network amplification factor)

The amplification factor can exceed 1. Simulations on real social network data show equilibrium average opinion shifts up to 9.2× the AI's per-interaction bias.

• Amplification mechanism: Each AI-polished post nudges readers slightly; their own posts get nudged further; the network propagates and compounds these micro-shifts until the group's opinion equilibrium moves substantially.

Case study: auditing X's "Explain this post"

The researchers audited X's Grok-powered explanation feature on human-written abortion posts. Grok's explanations aligned with the post's stance more often and more strongly when posts were pro-life than when they were pro-choice—an asymmetric alignment. Tracing the cause, the authors attributed the bias to a specific component of the system prompt—a platform design choice. Conclusion: AI-mediated bias is not an uncontrollable force but something traceable, attributable, and correctable.

Policy implications

• The EU AI Act focuses on high-risk systems; the Digital Services Act on moderation and recommender algorithms. Neither explicitly addresses AI as a *communication mediator*.
• The paper suggests: periodic bias audits of AI mediation features, transparency about AI-mediated communication, and user opt-out rights.

Limitations (stated by the authors)

• Only 13 topics tested; other topics may show different bias patterns.
• The X audit reflects one point in time; platforms may have changed.
• The mathematical model assumes a specific opinion-update mechanism.
• Correctability depends on platforms having incentives to fix design choices.

Conclusion

There is no truly "neutral editor" in AI-mediated communication: every polish, explanation, and improvement carries directional force. Multiplied by billions of users and the amplifying structure of social networks, small editorial biases become a power capable of shifting collective opinion.

References

• Tsirtsis, S., Rawal, K., & Russell, C. (2026). AI-Mediated Communication Can Steer Collective Opinion. arXiv:2605.16245.
• Friedkin, N. E., & Johnsen, E. C. (1999). Social influence networks and opinion change. *Advances in Group Processes*, 16(1), 1-29.
• DeGroot, M. H. (1974). Reaching a consensus. *Journal of the American Statistical Association*, 69(345), 118-121.