The Impact of AI-Mediated Communication on Collective Opinions: The Amplification Effect of Platform Algorithmic Bias

Through empirical analysis, theoretical modeling, and platform audits, this article reveals core conclusions: When large language models (LLMs) act as intermediaries in human communication, they introduce directional biases in edited text; these biases are significantly amplified through social network propagation, driving collective opinions to shift in specific directions. Taking X platform's Grok feature as an example, platform design choices (such as training data, system prompts) lead to specific biases (e.g., pro-life tendencies on the abortion issue). The study provides a scientific basis for the governance and policy formulation of AI-mediated communication.

Prevalence of AI-Mediated Communication

Generative AI is deeply integrating into online communication:

• Content generation assistance: Polishing social media posts
• Content explanation and summarization: Providing explanation functions for controversial content
• Real-time translation and transcription: A bridge for cross-language communication
• Recommendation systems: Indirectly shaping the discussion environment

Research Gaps

Existing studies have confirmed the biased impact of AI on individual opinions, but the impact of AI-mediated communication on collective opinion formation has not been fully explored: Are individual-level biases amplified through network effects? Do tiny AI biases lead to significant shifts in collective opinions? These questions are crucial for understanding AI's social role.

Empirical Experiment Design

• Multi-model coverage: Ensuring the universality of findings
• Controversial topics: Social divisive issues such as gun control and atheism
• Task: Asking LLMs to polish/explain user text

Theoretical Model Construction

Treat social networks as graph structures and add an AI mediation layer:

• Nodes: Individual users
• Edges: User interaction connections
• Core mechanisms: Expression transformation (AI modifies published content), perception transformation (AI explanations influence recipients), network propagation

Simulation and Audit Methods

• Simulation: Using real social network topology data, initializing diverse opinion distributions, and introducing AI biases
• Platform audit: Collecting posts with different positions on X's Grok feature (abortion issue) and analyzing the biases in explanations

Empirical Findings

LLMs have systematic directional biases when editing text:

• Gun control: Tend to support regulation
• Atheism: Tend to oppose atheism

Theoretical and Simulation Results

• Bias amplification effect: Tiny AI biases are significantly amplified through network propagation
• Collective opinion shift: The average opinion of the network moves toward the direction of AI bias
• Heterogeneous impact: Central nodes are more affected

Platform Audit Results

Grok has a pro-life bias in explanations on the abortion issue:

• Language choice: Using pro-life terminology
• Information emphasis: Highlighting pro-life arguments
• Root cause: Platform design choices (training data, system prompts, safety policies)

Transparency Requirements

• Bias disclosure: Platforms need to publicly disclose the types and directions of AI mediation biases
• Editing notifications: Clearly inform users when AI modifies content
• Algorithmic audits: Regular third-party bias audits

User Control

• Selective use: Users can choose whether to enable AI mediation
• Bias adjustment: Provide options to adjust the degree of bias
• Access to original content: Allow viewing of unmodified original content

Platform Responsibility

• Design responsibility: Be responsible for design choices such as training data and system prompts
• Impact assessment: Conduct social impact assessments before deployment
• Continuous monitoring: Establish a monitoring mechanism for the impact on collective opinions

Research Limitations

• Topic coverage: Focused on a few social controversial topics; other fields (economy, tech ethics) need exploration
• Cultural background: Based on Western platforms; non-Western cultural contexts need to be studied
• Long-term effects: Simulations and audits reflect short-term effects; long-term social impacts are not fully captured

Future Research Directions

• Expand to multilingual and multicultural contexts
• Develop real-time monitoring tools for AI mediation impacts
• Explore technical solutions for bias mitigation
• Study the moderating effect of user awareness on bias effects