"It's Complicated": Negotiating Accessibility and (Mis)Representation in Image Descriptions of Race, Gender, and Disability

Document Title

“It’s Complicated”: Negotiating Accessibility and (Mis)Representation in Image Descriptions of Race, Gender, and Disability

Document Information

• Subject Area: Human-Computer Interaction (HCI), Accessibility Studies, Representations of Race, Gender, and Disability
• Keywords: Accessibility, AI, Blind Users, Image Descriptions, Disability, Gender, Race, Visual Impairment

Research Background and Issues

• Existing Problems or Challenges:
  1. Current image description guidelines (e.g., Web Content Accessibility Guidelines, WCAG) provide general advice but lack specific regulations on describing the appearance of individuals in photos.
  2. Misrepresentation in descriptions of race, gender, and disability can lead to discrimination, bias, or unfair stereotypes.
  3. While AI-generated image descriptions may alleviate some accessibility issues, they risk propagating biases and amplifying discriminatory errors.
• Importance of the Issue:
  1. For blind users who rely on screen readers, image descriptions are a critical channel for accessing visual information.
  2. The way descriptions are framed directly impacts users’ understanding of individuals in images and can have significant political and sociocultural implications.
  3. Inappropriate or automated generation of descriptions may exacerbate injustices faced by marginalized groups.
• Research Motivation and Related Work:
  1. The study explores how blind users address issues of race, gender, and disability representation through non-visual means.
  2. While there is preliminary research on AI bias and accessibility, studies specifically addressing the description of individuals’ appearances in images are scarce.

Solutions

• Primary Methods and Solutions:
  1. Conducted interviews with 25 blind screen reader users, including Black, Indigenous, People of Color (BIPOC), non-binary, and/or transgender individuals, to explore their perspectives and preferences regarding image descriptions and appearance representation.
  2. Proposed ethical principles to guide the description of appearance information related to race, gender, and disability.
  3. Analyzed the strengths and weaknesses of AI-generated descriptions and the potential benefits and harms of AI technology for these groups.
• Innovative Aspects:
  1. The study focuses on users with intersecting marginalized identities (e.g., race + gender + disability), offering a multidimensional perspective and addressing complex intersectional issues.
  2. Examines the limitations of AI in describing human appearances and its potential ethical risks.
  3. Expands the role of image descriptions from mere information transmission to cultural and identity negotiation.
• Implementation Steps and Key Techniques:
  1. In-depth interviews designed for specific user groups, covering topics such as self-presentation, misrepresentation, image description preferences, and AI usage experiences.
  2. Systematic analysis of language sensitivity and descriptive conflicts based on interview results.
  3. Proposed concrete recommendations: distinguishing appearance descriptions (e.g., skin tone) from assumed identities (e.g., racial labels).

Research Findings

• Main Findings:
  1. Participants expressed a desire for more detailed appearance information in image descriptions but emphasized the importance of distinguishing between descriptions and identity assumptions.
  2. Requirements for descriptive language included avoiding presumptive identity labels, respecting the preferences of those being described, and using specific appearance details.
  3. While AI-generated image descriptions were seen as a potential tool, participants expressed concerns about their accuracy and ethical implications, particularly regarding racial and gender biases.
• Advantages Over Existing Solutions:
  1. Provides specific guidelines to make image descriptions more practical and respectful of individuals’ identities.
  2. Compared to existing research, offers a more comprehensive consideration of user needs, including linguistic and cultural nuances in image descriptions.
• Experimental or Evaluation Results:
  1. In the practice of describing images on social media, verbs (e.g., actions), clothing, and background elements were commonly included, while race, gender, or disability descriptions were less frequent.
  2. Although AI-generated descriptions offered some novel uses, participants noted that biases in AI could cause significant harm (e.g., misclassifying gender or age).
• Limitations and Future Directions:
  1. Since the participant sample leaned toward individuals open to discussing their identities, the findings may not fully represent the experiences of users who prefer not to disclose their identities or backgrounds.
  2. The study focuses on user needs rather than providing more specific support for AI technologies, such as developing content prototypes for AI-generated descriptions.
  3. Recommends further exploration of workflows for professional image describers and content creators to address accessibility gaps caused by missing descriptions.

Output Summary

This paper provides detailed insights into accessible image descriptions for blind users with intersecting marginalized identities. By emphasizing careful use of descriptive language, proposing potential improvements to AI technology, and offering practical recommendations to enhance user experiences while minimizing harm, the article calls for deeper exploration of the boundaries between accessibility research and AI ethics.