Labeling Synthetic Content: User Perceptions of Label Designs for AI-Generated Content on Social Media

Research Background and Issues

• What problems or challenges did the authors identify?

  • The rapid dissemination of deepfake content or AI-modified content generated by generative AI on social media threatens the transparency and authenticity of the digital environment.
  • Users often question the authenticity of AI-labeled content and lack effective means to determine whether content has been AI-generated or edited.
  • Existing misinformation label designs lack specific focus on AI-generated content and fail to fully consider user behavior and the practical application scenarios of labels.

• Why is this issue important?

  • As AI-generated content technology rapidly evolves, transparent labeling can mitigate the spread of misinformation and misleading behavior, enhancing public trust in social platforms.
  • Labels can help users make more informed interaction choices, reducing the spread of misleading content, which is particularly important for protecting democracy and reducing social division.

• Research Motivation and Related Work

  • This study is driven by the European AI Act and U.S. legal implementations, which call for proactive labeling of deepfake images and similar content to enhance transparency.
  • Previous studies on misinformation warning label design have demonstrated their effectiveness in reducing user belief in misleading content, but there is a lack of systematic evaluation of warning label designs specifically for AI-generated content.

Solutions

• What methods or solutions did the authors propose?

  • A label design framework comprising four main design dimensions (label sentiment, icons and colors, label placement, and label detail level) was proposed.
  • Based on the framework, 10 warning label samples were developed to evaluate their effectiveness in labeling AI-generated content.
  • Experimental setup: 911 participants were randomly assigned to a control group (no label design) and 10 treatment groups (different label designs) to test trust in labels, perceived authenticity of content, and social media interaction behavior.

• What are the innovations of this solution?

  • The focus of AI-generated content labels is on improving transparency, reducing the risk of misinformation, and building user trust in social media platforms and the labels themselves.
  • A systematic design space was constructed, and various aspects of user perception of labels were experimentally evaluated, providing empirical support for designing more effective labeling systems.

• What are the implementation steps? What key technologies were used?

  • Step 1: Define the four dimensions and options of the label design space based on existing literature and social media platform label designs.
  • Step 2: Derive 10 label samples from the design space and embed them into different types of social media images (entertainment and political content).
  • Step 3: Experimentally evaluate user trust in labels (Trust-In-Label), trust in platforms (Trust-In-Platform), perceived authenticity of content (Belief-In-Content), and interaction behaviors with social media content (Like, Comment, Share).

Research Findings

• What specific findings were achieved?

  • Impact of labels on perceived content authenticity: All label designs significantly enhanced users' belief that the content was AI-generated or modified.
  • Differences in trust based on label design: Trust in labels varied significantly by design (especially label language and icons). For example, the "Content Credentials" label with detailed information demonstrated the highest trust, though its information density might increase cognitive load for users.
  • Impact on trust in social media platforms: Trust in labels was positively correlated with trust in social media platforms.
  • User interaction behavior: Labels had no significant impact on users' "like," "comment," and "share" behaviors, but political content was more influenced than entertainment content.

• What advantages does it have over existing solutions?

  • Provides preliminary empirical support, analyzing how warning label design affects user behavior and psychology.
  • Considers the effects of labels on different types of content (political vs. entertainment), offering targeted design recommendations applicable to diverse scenarios.

• What are the experimental or evaluation results?

  • Label designs using "Made with AI" language (e.g., samples 6 and 9) were particularly effective in enhancing perceived content authenticity.
  • Label placement significantly influenced user experience; for instance, labels placed above or below the image were more readily accepted.
  • Despite the presence of labels, changes in user interaction behavior remained moderate, indicating that labels had limited overall impact on social behaviors.

• Limitations and Future Directions

  • Limitations:
    • The effect sizes of the experimental results were small, suggesting limited practical impact on user behavior.
    • The controlled experimental environment may not fully reflect the complexity of real-world user behavior.
    • The label design samples exhibited bias and did not sufficiently cover a broader range of design options.
  • Future Directions:
    • Test more label variants, especially watermark-based or hybrid label designs.
    • Validate the long-term effectiveness of label designs and changes in user behavior in real-world social network scenarios.
    • Conduct future studies targeting non-technical user groups and diverse global regions to explore cultural differences in label acceptance and effectiveness.

Through this work, the article provides an early framework and empirical support for designing transparent and effective labels for AI-generated content, emphasizing the importance of label language, appearance, and placement in shaping user perception and trust. These research findings offer significant reference value for industry practices and related policy development.