Interface Support for Evaluating Disability Bias in AI Generated Images

Paper Title

Interface Support for Evaluating Disability Bias in AI-Generated Images

Publication Info

• Topic area: Addressing disability bias in AI-generated text-to-image outputs through user-facing interventions.
• Keywords: AI bias, text-to-image models, disability representation, stereotypes, user interfaces, education intervention, AI feedback, generative AI, prompt engineering, disability studies.

Background and Problem

• Problem / challenge: Generative text-to-image (T2I) models often replicate stereotypes and biases about disabled people, producing inaccurate or harmful representations. Existing strategies like dataset improvements and model fine-tuning have not fully mitigated these biases.
• Significance: Addressing disability bias is critical for ensuring equitable AI systems, particularly as T2I models are increasingly used in diverse applications like education, marketing, and art.
• Motivation and related work: Prior research has documented biases in AI-generated images, such as "inspiration porn" and exaggerated portrayals of disability. While end-user auditing has shown promise in identifying bias, users often lack the expertise to assess disability stereotypes effectively. This paper explores whether user-facing interventions can empower non-expert users to identify and avoid biased representations.

Solution

• Proposed approach: Development of two user-facing interventions: (1) an education module explaining disability stereotypes and (2) AI-generated feedback analyzing images for stereotypes.
• Novelty:
  1. Empirical evaluation of two interventions to support stereotype detection in AI-generated images.
  2. Analysis of user preferences for disability representation in images.
  3. Design implications for improving AI interfaces and prompt engineering.
• Procedure and key techniques:
  • Education intervention: A one-page module describing four stereotype categories (pity, extraordinary, medical/mortality, inaccurate assistive technologies) with examples and guidelines for better representation.
  • AI feedback intervention: Real-time analysis of images using ChatGPT (gpt-4o-mini) to detect stereotypes and provide brief explanations.
  • Evaluation: Controlled experiment (N = 103) measuring changes in image ratings pre- and post-intervention, and qualitative study (N = 10) exploring user experiences with both interventions.

Results

• Concrete findings:
  • The Education intervention significantly reduced participants' likelihood of using images with stereotypes (average rating drop: 0.6 points).
  • AI Feedback intervention showed no statistically significant effect but was rated as helpful by participants.
  • AI feedback accuracy averaged 80%, with a 34% false-negative rate and 4% false-positive rate.
• Advantage over baselines:
  • Participants exposed to the Education intervention were 85% more likely to lower their ratings for stereotypical images compared to those without the intervention.
  • AI Feedback intervention highlighted stereotypes but suffered from over-reliance by users, who accepted incorrect feedback more than 50% of the time.
• Experiments / evaluation:
  • Controlled experiment (N = 103): Participants rated images pre- and post-intervention, assessing representation quality and likelihood of use.
  • Qualitative study (N = 10): Participants used a prototype to generate images and provided feedback on interventions and prompting challenges.
  • Metrics: Likert scale ratings, thematic analysis of open-ended responses, and accuracy comparison of AI feedback against expert-coded ground truth.
• Limitations and future work:
  • Limited representation of diverse demographics in the participant pool.
  • AI Feedback intervention was error-prone, with substantial over-reliance by users.
  • Future work should explore co-design with disabled communities, culturally responsive stereotype definitions, and nonvisual interface adaptations.

Summary

This paper investigates user-facing interventions to address disability bias in AI-generated images. The Education module significantly reduced the likelihood of using stereotypical images, while the AI Feedback intervention showed mixed results due to accuracy limitations and user over-reliance. Participants prioritized realistic images where subjects "looked disabled" but varied in their preferences for tone, with some favoring everyday representations and others emphasizing struggle or suffering. Findings highlight opportunities for improving AI interfaces, supporting prompt engineering, and aligning outputs with disabled community values.