Technical Summary: AI-Mediated Communication Can Steer Collective Opinion

Problem Statement

Generative Artificial Intelligence (AI), particularly Large Language Models (LLMs), is increasingly integrated into online platforms to mediate human-to-human communication. Features such as "Improve my post" on LinkedIn or "Explain this post" on X modify user-generated content or provide context for it. While prior research has established that LLMs can express biased opinions and influence individuals during direct human-AI interactions, there is a significant gap in understanding how these biases affect collective opinion formation when AI acts as a mediator between humans. The central question addressed is: How do biases introduced by an AI system in human-to-human communication propagate through a social network to shift the collective opinion of a population?

Methodology

The authors employ a tripartite methodology combining empirical analysis, theoretical modeling, and platform auditing.

1. Empirical Analysis of Directional Bias

The study investigates whether LLMs introduce directional biases when editing or drafting social media posts on contested topics.

• Datasets: The authors utilized the UKP Sentential Argument Mining Corpus (8 topics) and the SemEval-2016 Task 6 Dataset (6 topics), covering issues like abortion, gun control, and atheism.
• Models: Four open-weight LLM families were tested: Llama-3.1-8B-Instruct, Ministral-3-8B-Instruct-2512, gemma-3-12b-it, and Qwen3-8B.
• Tasks: Two scenarios were emulated: (i) drafting a post from an argument (argument $\to$ post) and (ii) improving an existing draft (post $\to$ post).
• Measurement: Opinions were quantified on a continuous scale $[0, 1]$ (0 = against, 1 = in favor) using an ensemble of five classifiers based on different pretrained text embedding models. The bias was measured as the difference ($\beta = y - x$) between the original human opinion ($x$) and the transformed LLM opinion ($y$).

2. Theoretical Modeling of Opinion Dynamics

To analyze the propagation of bias, the authors extended the seminal Friedkin-Johnsen model of opinion dynamics.

• Model Structure: A social network is modeled as a weighted graph where individuals update their opinions based on innate beliefs and the perceived opinions of neighbors.
• AI Transformation: A function $f$ is introduced to represent the AI mediation. The perceived opinion of a neighbor $j$ is transformed from their expressed opinion $x_j$ to $y_j = f(x_j)$.
• Linear Analysis: The authors first analyzed a linear transformation $f_{lin}(x) = mx + b$. They derived closed-form expressions for the equilibrium state and proved that the system converges to a unique equilibrium.
• Key Insight: The model treats the AI as an "invisible neighbor" with a persistent neutral point $\nu$. The authors derived conditions under which the shift in the population's average opinion at equilibrium strictly exceeds the AI's initial "one-off" bias introduced to individual inputs.

3. Simulation and Platform Auditing

• Simulations: Using real social network data from the SNAP repository (Twitter, Facebook, Google+), the authors simulated opinion dynamics over 100 time steps. They applied non-linear AI transformations estimated from their empirical results (using Nadaraya–Watson kernel regression) to a fraction of users ("AI adopters").
• Platform Audit: The authors audited X's "Explain this post" feature, which uses the Grok model. They analyzed 78 abortion-related posts (39 pro-choice, 39 pro-life) to determine if the AI-generated context aligned with specific stances. They systematically ablated X's four prompt guidelines to identify the source of any observed bias.

Key Results

1. Systematic Directional Biases in LLMs

Empirical results demonstrate that LLMs systematically introduce directional biases when editing human text, even when instructed to preserve the original meaning.

• Magnitude and Direction: Most models (Llama, Ministral, gemma) introduced a statistically significant bias toward the "in favor" position across most topics (e.g., gun control, feminism, climate change).
• Exceptions: On the topic of atheism, models tended to introduce a bias against the "in favor" position, despite often expressing a positive opinion when asked directly. This suggests that benchmarks measuring direct LLM opinions are insufficient to capture biases introduced during mediation.
• Correlation: There is a moderate positive correlation between an LLM's directly expressed opinion and the bias it introduces, suggesting the model's inherent training data "leaks" into its editing capabilities.

2. Amplification of Bias in Social Networks

Theoretical and simulation results confirm that AI-mediated biases are not merely additive but are amplified through the network.

• Equilibrium Shift: In the linear model, the shift in the population's average opinion at equilibrium can be significantly larger than the average one-off bias introduced by the AI.
• Simulation Findings: In simulations using the Twitter network, the long-run average opinion shift was up to 9.2 times larger than the average bias introduced to individual innate opinions. The amplification effect is strongest when the user population has balanced innate opinions but low stubbornness (high susceptibility to influence).

3. Design-Induced Bias on X

The audit of X's "Explain this post" feature revealed a directional bias in Grok's outputs regarding abortion.

• Asymmetry: Grok was significantly more likely to generate context supporting a pro-life stance when explaining a pro-life post compared to generating pro-choice context for a pro-choice post.
• Root Cause: By ablation testing the prompt guidelines, the authors identified that a specific guideline—"Provide truthful and based insights, challenging mainstream narratives if necessary, but remain objective"—was the primary driver of this pro-life bias. Removing this guideline eliminated the statistically significant bias.

Significance and Claims

The paper argues that AI-mediated communication represents a novel and potent lever for influencing collective opinion, with implications that current regulatory frameworks may fail to address.

• Amplification Mechanism: The core theoretical contribution is the demonstration that small, individual-level biases introduced by AI can compound and propagate through social networks, leading to large-scale shifts in collective opinion that far exceed the initial bias magnitude.
• Design Responsibility: The audit of X demonstrates that these biases are not inevitable artifacts of model training but can be directly traced to specific platform design choices (prompt engineering). This suggests that platforms have the agency to shape opinion formation through their interface design.
• Regulatory Gaps: The authors contend that existing EU legislation, specifically the AI Act (AIA) and the Digital Services Act (DSA), may be insufficient to mitigate these risks.
  • The AIA's "high-risk" classification for election influence may not apply if the AI's explicit goal is not to influence voting, even if the effect is to shift opinions.
  • The DSA's systemic risk definitions regarding misinformation or human rights violations may not capture "biased outputs" that are not factually false but are directionally skewed.
• Broader Implications: The authors warn that if journalists, professionals, and the public increasingly rely on AI to draft, refine, or contextualize information, these systems could silently shape the text people read and write, potentially eroding the diversity of collective knowledge and democratic discourse.

The paper concludes by calling for further research into the causal links between AI mediation and voting behavior, and for policymakers to consider the subtle, cumulative effects of AI on opinion formation when crafting future regulations.