AI-Driven Mediation Strategies for Audience Depolarisation in Online Debates

Title of the Paper

AI-Driven Mediation Strategies for Audience Depolarisation in Online Debates

Paper Information

• Subject Area: Human-Computer Interaction (HCI), Psychology, Applications of Artificial Intelligence in Social Media Debates
• Keywords: Social media, debates, mediation, artificial intelligence, depolarisation, generative AI, psychology, human-computer collaboration, chatbots

Research Background and Problem

• Identified Problems or Challenges:

  • Polarisation on social media exacerbates societal divisions, and simple content moderation methods (e.g., deleting or flagging controversial content) fail to address the deeper disagreements underlying debates.
  • Current content moderation approaches suppress open discussions and may shift controversies to anonymous forums and platforms with more severe echo chamber effects.
  • Audiences (non-participants in debates) are a potential key demographic for depolarisation efforts on social media, yet these audiences are susceptible to non-constructive viewpoints on divisive issues.

• Importance:

  • Overly polarised discussions not only affect participants but also negatively impact passive audiences, leading to feelings of helplessness regarding social issues (existential dread) and further erosion of social trust.
  • Compared to simple content moderation, fostering in-depth discussions and proposing innovative solutions can strengthen public critical thinking and reduce tendencies toward extremism.

• Research Motivation and Related Work:

  • By integrating psychological conflict mediation theories with generative language models, the focus is placed on non-coercive and constructive information mediation.
  • This paper references the Thomas-Kilmann Conflict Mode Instrument (TKI) and its five conflict resolution strategies to design AI mediation strategies.
  • Current research predominantly focuses on content moderation and bias detection, with limited exploration of psychology-driven AI mediation frameworks.

Solution

• Proposed Method or Solution:

  • Utilizing the GPT-4 generative language model to design mediation bots based on TKI’s five mediation strategies (collaborating, compromising, competing, accommodating, avoiding).
  • Mediation bots intervene in debates through three approaches: “questioning, statements, and responses,” aiming to provide new perspectives or consensus-driven solutions.
  • Applying mediation strategies in simulated online debate environments and analyzing their effects on audience depolarisation and perceptions of the debate.

• Innovative Aspects:

  • The first complete application of psychology’s TKI conflict mode instrument in AI mediation design.
  • Fine-tuning large language models (prompt tuning) to generate mediation content consistent with specific psychological strategies, extending psychological mediation theories and enabling real-world applications on social platforms.
  • The research focuses primarily on debate audiences rather than participants, aiming to influence a broader range of social media users.

• Implementation Steps and Key Technologies:

  1. Generating Mediation Content:
     • Using GPT-4 to implement TKI-based mediation strategies.
     • Ensuring standardization and consistency in mediation phrasing (e.g., unified templates for questions and responses).
  2. Experimental Design:
     • Selecting five controversial and diverse debate topics (e.g., mandatory COVID-19 vaccination, acceptance of Russian-Ukrainian refugees).
     • Designing Reddit-style visual posts to simulate real user debate scenarios.
  3. Research Questions (RQs):
     • RQ1: The impact of mediation strategies on audience perception of persuasiveness.
     • RQ2: The effect of mediation strategies on audience opinion changes and depolarisation.
     • RQ3: How mediation strategies influence audience perceptions of the potential for consensus in debates.
  4. Quantitative and Qualitative Evaluation:
     • Using the Perceived Argument Strength (PAS) metric to score the persuasive strength of mediation content.
     • Monitoring audience opinion changes and perceptions of consensus potential via 7-point Likert scales.

Research Findings

• Specific Results:

  • Effectiveness of Mediation Strategies:
    • Highly collaborative strategies (e.g., collaborating and accommodating) were more effective in enhancing argument persuasiveness (PAS) and promoting opinion depolarisation.
    • The “accommodating” strategy performed the best, achieving depolarisation effects three times greater than scenarios without mediation.
    • The “competing” strategy, while assertive, exacerbated opinion divergence.
    • The “compromising” strategy unexpectedly led to further polarisation among some respondents.
  • Psychological Impact on Audiences:
    • Highly collaborative mediation strategies were more effective in eliciting critical thinking, encouraging audiences to consider both sides of the debate more comprehensively.
    • Some low-collaboration or avoidance strategies also reduced direct confrontation by shifting the topic.
  • Consensus Perception:
    • Mediation bots employing the “compromising” strategy significantly improved audience evaluations of the likelihood of achieving consensus between debate participants.

• Advantages Compared to Existing Solutions:

  • Unlike traditional content moderation, the proposed mediation strategies do not censor speech but promote discussion by presenting constructive viewpoints.
  • By focusing on audiences rather than debate participants, the mediation’s impact extends from individuals to a broader social media user base.

• Experimental or Evaluation Results:

  • Analysis of 144 audience members confirmed that the fine-tuned GPT-4 could generate highly consistent psychology-driven mediation outputs (88.21% strategy match rate).
  • Data indicated that highly collaborative mediation strategies were most effective in depolarisation and enhancing audience perceptions of consensus potential.

• Limitations and Future Directions:

  • This study focuses solely on audience responses to mediation content, with insufficient exploration of dynamic reactions from debate participants and adaptive bot strategies.
  • Future research should investigate customized mediation for different community languages and cultures, as well as trigger mechanisms combining mediation and content moderation.
  • Ethical considerations regarding language models, such as maintaining neutrality and preventing potential biases, also warrant further study.