Public Perceptions About Emotion AI Use Across Contexts in the United States

Research Background and Issues

Issues and Challenges

• The application scope of Emotion AI is continuously expanding, from workplaces to healthcare, from smart devices to public safety. However, public attitudes and comfort levels regarding its use in these areas remain largely unknown.
• The core challenge lies in the lack of systematic research on public perceptions and acceptance of Emotion AI across different scenarios, particularly regarding variations based on identity and context (e.g., race, gender, disability status).

Significance of the Research

• The deployment of Emotion AI technologies could significantly impact issues such as privacy, algorithmic bias, and emotional labor. Certain groups (e.g., people of color, individuals with disabilities, and gender minorities) may be more affected than others.
• The absence of quantitative research on public attitudes creates challenges for policy-making, especially as the U.S. currently has minimal regulation on Emotion AI and emotional data.

Related Work and Motivation

• Previous studies have focused on qualitative investigations of Emotion AI in a few specific domains (e.g., social media, workplaces) but lack large-scale quantitative research.
• Similar surveys in the UK found general public disapproval of Emotion AI but did not explore its prevalence across contexts or the influence of identity factors.
• Motivation for this study: To provide the first cross-sectional survey of U.S. public attitudes toward Emotion AI, with a core focus on the interaction between context, emotional types, and identity factors.

Solution

Methods and Process

• Data Collection: Recruited 599 participants via the Prolific survey platform, based on U.S. population proportions (gender, age, race, and political orientation), with oversampling of minority identity groups (individuals with disabilities, people of color, transgender, and non-binary individuals).
• Survey Design:
  • Provided a standard definition of Emotion AI and its potential tasks (e.g., inferring emotions through voice or facial expressions).
  • Measured public attitudes toward Emotion AI overall and comfort levels in 11 high-impact domains (e.g., healthcare, border control, public spaces).
  • Assessed comfort with specific emotional inferences (e.g., happiness, anger, sadness).
  • Controlled variables included race, gender, age, education, income, and perceived AI accuracy.
• Analysis Methods: Used Type II ANCOVA to measure the significant impact of identity factors on comfort levels, with post-hoc Tukey HSD tests to explore significant contrasts.

Research Innovation

• Quantified, for the first time, public attitudes and comfort levels toward Emotion AI in the U.S., spanning 11 usage scenarios and detailing identity characteristics.
• Highlighted the significant impact of specific emotional types (e.g., happiness, fear) on public comfort levels.
• Conducted cross-analysis to explore the interactive effects of identity factors on public perceptions, offering direct insights for the equitable deployment of Emotion AI.

Research Findings

Key Findings

1. Overall Attitudes: Participants expressed generally negative attitudes toward Emotion AI (mean=3.30/7).
2. Cross-Context Differences:
   • Comfort levels were highest in healthcare contexts (mean=3.61/7) and lowest in workplace and social media contexts (means of 2.34 and 2.38, respectively).
3. Impact of Emotional Types:
   • Participants were significantly more comfortable with Emotion AI inferring "happiness"; "fear" was perceived as having some benefits in certain safety-related contexts.
4. Impact of Identity Factors:
   • Participants with disabilities expressed significantly greater discomfort with Emotion AI use in various contexts (e.g., workplaces, vehicles, public spaces).
   • People of color showed higher acceptance of Emotion AI in most contexts but expressed greater discomfort in scenarios with potential racial bias risks, such as public spaces, border control, and job interviews.
   • Significant gender differences were observed: men were generally more accepting of Emotion AI than women, while transgender and non-binary individuals were particularly uncomfortable with its use in public and workplace settings.

Comparison with Existing Research and Advantages

• Compared to UK studies, this research expanded to include identity dimensions such as gender and disability, revealing differences in public attitudes between the U.S. and the UK (e.g., greater emphasis on privacy and racial factors in the U.S.).
• Provided more granular integrated analysis by combining contextual factors, specific emotional inferences, and identity intersections.

Experimental or Evaluation Results

• Low overall comfort with Emotion AI reflects widespread privacy concerns and skepticism about its accuracy and fairness.
• AI inferences of "happiness" and those perceived as "accurate" had a positive effect on public comfort levels, but the overall increase was limited.

Limitations and Future Directions

• Limitations:
  • Did not differentiate between subtypes of disabilities or racial groups in their impact on comfort levels.
  • The effect sizes between certain contexts and emotional types were small, requiring cautious interpretation.
• Future Directions:
  • Further exploration of specific risks and benefits in particular contexts (e.g., education and healthcare).
  • Expanding the survey to other countries or cultural backgrounds to compare cross-national differences in attitudes.
  • Investigating how policies and design can more effectively address the concerns of minority groups.

Conclusion

This study uses comprehensive quantitative data to validate the complex public attitudes toward Emotion AI across contexts and identity intersections, highlighting public expectations for the protection of emotional data. The findings directly call for stricter cross-context regulation of Emotion AI by regulatory agencies and provide actionable ethical guidelines for designers and developers.