Generative AI and Perceptual Harms: Who’s Suspected of using LLMs?

Research Background and Problem

• Issues and Challenges: The authors identify that when text generated by large language models (LLMs) is suspected of being AI-produced, audiences may negatively evaluate its quality and the user behind it. This phenomenon, termed "perceptual harms," could disproportionately affect certain social groups.
• Significance: LLMs are rapidly being adopted in contexts such as work and education. While these technologies facilitate the creation of high-quality content, they may also harm certain social groups, exacerbating existing inequities faced by historically marginalized populations.
• Research Motivation: Previous studies have shown that when text is suspected to be AI-generated, people tend to lower their evaluations of the content and their trust in the author. However, these studies have not explored how such harms vary across different social groups. This paper aims to fill this gap by uncovering new forms of social bias within AI technologies.

Solution

• Approach and Framework: The authors introduce the concept of "perceptual harms" and assess its impact through online experiments. These harms occur when individuals suspected of using AI face differential social treatment, regardless of whether AI was actually used.
• Experimental Design:
  • Three experiments focus on gender (male vs. female), race (White vs. Black), and nationality (U.S. vs. East Asian).
  • The experiments simulate an online freelance marketplace where participants evaluate whether content created by professionals from different groups was AI-generated, rate its quality, and make hiring decisions.
  • Using a randomized design, participants are presented with fictional resumes from various groups, containing both human-written text and text modified to exhibit "AI-like" characteristics.
• Key Techniques: The study employs trained models to generate fictional avatars and content, manipulating elements such as linguistic style to test participants' responses to perceptual harms.

Research Findings

• Specific Results:
  • There are indeed differences in the suspicion of AI use across social groups. For instance, individuals from male and East Asian backgrounds are more likely to be suspected of using AI, while those from female and non-U.S. backgrounds are less likely to face such suspicion.
  • Regardless of social group, suspicion of AI use significantly reduces evaluations of text quality and hiring decisions.
• Advantages Over Existing Solutions:
  • This study is the first to define and systematically evaluate "perceptual harms," addressing not just technical issues but also unveiling new harms driven by social biases.
• Limitations and Future Directions:
  • Limitations:
    • Cultural bias exists, as the experimental results are based on U.S. participants and may not generalize to other regions.
    • The simulated scenario (online hiring marketplace) may not fully capture the complexity of real-world decision-making.
    • Participants might have been influenced by the experiment's focus on the theme of "AI use."
  • Future Directions:
    • Expand experiments to different cultural contexts and domains, such as art, journalism, and education.
    • Explore design interventions (e.g., AI usage disclosure labels or transparency mechanisms) to mitigate perceptual harms.
    • Investigate whether these harms diminish or evolve as societal acceptance of AI increases.

Conclusion

This paper introduces the concept of "perceptual harms" and validates its existence through experiments, addressing a gap in the field. The study demonstrates that suspicion of AI use affects evaluations of content quality and professional opportunities, with such effects potentially exacerbating the disadvantages faced by historically marginalized groups. Future work should focus on reducing these harms through social and technical interventions and exploring how the dynamic evolution of AI impacts perceptual harms.