Understanding the Effects of AI-based Credibility Indicators When People Are Influenced By Both Peers and Experts

Research Background and Issues

What issues or challenges did the authors identify?

1. Proliferation of online misinformation: Online social media is rife with false information, leading to severe consequences for public opinion, political stability, and health-related decision-making.
2. Accurately navigating complex social environments: Users are influenced not only by peer groups (ordinary users) but also by experts, and the interaction between the two further complicates information processing.
3. Uncertainty about the effectiveness of AI credibility indicators: Although artificial intelligence (AI) has been employed to label content credibility, its actual effectiveness in enhancing users' ability to identify misinformation, especially under social influence, remains unclear.

Why is this issue important?

• The spread of misinformation can lead to misunderstandings, panic, and even threats to life (e.g., misinterpretation of health information).
• Existing methods for filtering misinformation (e.g., manual fact-checking) are not scalable, and AI can help improve efficiency and reach.
• Understanding human reliance on AI credibility indicators under complex social influences is crucial for developing more effective information intervention tools.

Research Motivation and Related Work

• Related Work: Previous studies have shown that AI credibility indicators, when displayed independently, can improve users' judgment. However, there is limited research on the effectiveness of AI models in socially influenced scenarios (e.g., environments with mixed peer and expert influences).
• Research Questions:
  • RQ1: Can AI credibility indicators help detect misinformation and reduce its spread under the influence of peers and experts?
  • RQ2: Does the consistency (or discrepancy) between experts and AI affect the effectiveness of AI credibility indicators?
  • RQ3: Does verifying experts' professional background alter the effectiveness of AI credibility indicators?
  • RQ4: How does the effectiveness of AI credibility indicators change when the accuracy of the AI model fluctuates?

Solutions

What methods or solutions did the authors propose?

• Experimental Design: Conduct three rounds of pre-registered randomized experiments, each with different variable controls (e.g., AI prediction accuracy, expert background) to observe the role of AI credibility indicators in various social influence scenarios.
• Research Metrics:
  • User judgment accuracy (frequency of correctly identifying true or false news).
  • Users' ability to distinguish true from false information (sensitivity).
  • Users' sharing intentions and their bias towards sharing true versus false information.

What are the innovative aspects of this solution?

1. Comprehensive social influence analysis: This study is the first to incorporate both peer and expert social influences into the research on AI credibility indicators.
2. Multi-factor consideration: It analyzes the dynamic impact of expert credibility (self-claimed vs. verified), AI model accuracy, and expert-AI consistency on user behavior.
3. Detailed experimental design: Through a two-stage experimental process (collecting "peer opinions" in advance and embedding expert judgments), the study simulates real-world social media interactions.

Implementation Steps and Key Techniques

1. Data Collection:
   • 40 health news articles (20 true and 20 false).
   • Stage 1: Gather judgments from initial participants to serve as "peer opinions."
   • Stage 2: Present these initial judgments to new participants, with AI credibility indicators provided in the experimental group.
2. Control of Experimental Variables:
   • AI accuracy: Varied at 100%, 80%, and 55%.
   • Expert background: Self-claimed expertise vs. platform-verified expertise.
   • Expert-AI consistency (agreement vs. conflict).
3. Data Analysis: Use binary logistic regression and mediation analysis to evaluate the direct and indirect effects of AI on user decision-making.

Research Findings

What specific findings were obtained?

1. Significant impact of AI on misinformation detection:
   • Across all experimental scenarios, AI credibility indicators improved users' accuracy and sensitivity in identifying false news.
   • AI's influence was particularly pronounced when it disagreed with expert opinions.
2. Changes in sharing behavior:
   • AI credibility indicators reduced users' inclination to share false news.
   • When AI and experts disagreed, users were more likely to rely on AI and less likely to blindly trust "inconsistent experts."
3. Indirect influence of AI on peer opinions:
   • AI predictions indirectly influenced users by affecting their reliance on peer judgments, reducing their dependence on peer group opinions.

What advantages does it have compared to existing solutions?

• Provides a comprehensive examination of complex social influences (simultaneous presence of peers and experts), whereas prior research often focused on individual or single influence sources (e.g., AI or standalone experts).
• Explores the potential risks of AI prediction errors and their systemic misleading effects on user groups.

Limitations and Future Directions

1. Simplified experimental scenarios:
   • Health news has clear "true" or "false" characteristics, but external validity may be affected in more complex or subjective domains (e.g., politics).
   • The roles and identities of peers and experts (e.g., credibility) were artificially assigned in the experiment, lacking the complexity of natural social network interactions.
2. Risks of AI errors:
   • When AI predictions are incorrect, users still struggle to discern errors, potentially leading to cascading misinformation.
3. Future Directions:
   • Incorporate multi-source credibility markers (e.g., combining expert consensus with AI).
   • Design methods to display prediction uncertainty to reduce users' blind reliance on AI.
   • Extend research to more complex social network environments and multi-domain news topics (e.g., issues of opinion polarization).

Through these three rounds of experiments, this study effectively reveals the opportunities and risks of AI credibility indicators under multiple social influences, providing important insights for technological optimization and policy development.