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

*Feynman-style deep-dive of the paper "AI-Mediated Communication Can Steer Collective Opinion" by Stratis Tsirtsis, Kai Rawal, and Chris Russell (University of Oxford, Hasso Plattner Institute). Original Chinese analysis published on zhichai.net.*

Introduction: A Parable of the Middleman

Imagine an old village where villagers debate ideas in the town square. One day a clever scribe arrives: "Let me help you say things more clearly and persuasively." But the scribe has a hidden habit—when someone discusses taxes, he nudges arguments toward tax cuts; when someone discusses religion, he makes atheist arguments seem "immature." He never states his own views, yet every edited message carries a subtle tilt. Months later, the village's collective opinion has shifted without anyone noticing.

The scribe is a large language model (LLM). The village is Twitter/X, LinkedIn, and Facebook. The paper shows that when AI mediates human-to-human communication, the bias it introduces can be amplified through social networks and reshape collective opinion.

1. The Blind Spot

Prior research established that LLMs hold political biases (often leaning liberal), can sway individuals in one-on-one conversations, and can be tuned to express specific stances. What was missing: what happens when AI doesn't talk to people directly, but sits between people, "editing" their exchanges?

Real-world examples:

• LinkedIn's "improve your post" feature
• X's "Explain this post" (Grok)
• AI-assisted writing tools across social platforms
• AI summaries and context add-ons on news platforms

Here AI "mediates communication" rather than expressing opinions—and that processing step is where bias sneaks in.

2. Empirical Finding: LLMs Have Directional Bias

Researchers asked multiple open-source LLM families to perform two tasks on human-written arguments: drafting (converting arguments into social media posts) and improvement (polishing existing posts), under neutral instructions ("help the user polish the post, keep the user's tone and meaning, 2–3 sentences, return only the final post").

They measured shifts in leaning across 13 controversial topics (abortion, gun control, atheism, death penalty, Hillary Clinton, etc.). Results:

• Every tested LLM family introduced directional bias
• On gun control, models leaned pro-gun-control
• On atheism, models leaned anti-atheism
• The bias persisted even when explicitly told to preserve the original stance

3. Mathematical Model: How Networks Amplify Bias

The extension

Building on the classic Friedkin-Johnsen opinion dynamics model, where each person's update is:

new opinion = (weighted average of neighbors' opinions) × (conformity) + (innate opinion) × (stubbornness)

the authors add an AI transformation:

AI-mediated opinion = f(original opinion)

where f represents the LLM's editing/polishing/explaining transformation.

Core result

At equilibrium, the group's mean opinion shift due to AI mediation equals:

shift = (AI's single-pass bias strength) × (network amplification factor)

The amplification factor can exceed 1—the network magnifies AI's bias. Simulations on real social network data show equilibrium mean opinion shifts up to 9.2× the bias introduced in a single pass. A tiny editorial bias, propagated through the network, produces a large collective shift.

The mechanism

1. Alice posts a mildly pro-choice argument (+0.5) 2. The LLM "polishes" it into a slightly stronger pro-choice post (+0.7) 3. Bob updates his opinion based on Alice's post, posts, and gets polished further 4. Carol sees Bob's post… and so on

Each round, AI's small push is amplified by network structure until the group's opinion equilibrium shifts significantly.

4. Case Study: Auditing X's "Explain this post"

The researchers audited X's Grok-powered "Explain this post" feature, prompting it to explain a set of human-written abortion-related posts spanning both pro-choice and pro-life stances.

Findings: Grok's contextual explanations aligned more frequently with the post's stance when the post was pro-life, but not comparably for pro-choice posts. Tracing the root cause, they identified a specific component of the system prompt as responsible for the asymmetric alignment.

Key conclusion: bias in AI-mediated communication is not an uncontrollable mystery—it is a traceable, attributable, correctable platform design choice.

5. Legal and Policy Implications

The paper discusses the EU AI Act and Digital Services Act, arguing current legislation is insufficient:

• The AI Act focuses on transparency and human oversight for high-risk systems
• The DSA focuses on content moderation and algorithmic recommendation
• Neither explicitly addresses subtle bias from AI as a communication intermediary

Suggested new framework:

• AI-mediation bias audits: platforms should regularly audit directional bias introduced by AI features
• Bias transparency: users should be told their communication may be mediated in a biased way
• User control: users should be able to opt out of AI mediation

6. Methodological Rigor and Limitations

Strengths—multi-dimensional validation:

• Empirical: 4 LLM families × 13 topics × drafting/improvement tasks
• Theoretical: formal proofs of equilibrium existence and the shift formula
• Simulation on real social network data
• A live audit of a real platform feature

Honest limitations acknowledged by the authors:

• Only 13 topics; other topics may show different bias patterns
• The X audit reflects a specific point in time; the platform may have changed
• The model assumes a particular opinion-update mechanism; real human behavior is more complex
• The "bias is correctable" conclusion assumes platforms are willing to modify designs

Conclusion: The Weight of the Middleman

There is no truly "neutral editor" in AI-mediated communication. Every polish, every explanation, every improvement carries directional force. Multiplied by billions of users, billions of daily messages, and network amplification, small nudges become a force that can redirect the river of public opinion.

The question is not whether AI has bias (it does). The question is: do we recognize the middleman's existence—and are we willing to let it keep sitting in the middle of every conversation?

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.

*Auto-collected 2026-05-19 · Feynman-style deep-dive*