Engaging Communities Meaningfully in Defining Disability Representation for AI Image Generation

Paper Title

Engaging Communities Meaningfully in Defining Disability Representation for AI Image Generation

Publication Info

• Topic area: Disability representation in AI-generated visual media.
• Keywords: Disability representation, AI image generation, human-centered AI, community engagement, assistive technology, inclusive datasets, AI evaluation, participatory design, marginalized communities, advocacy organizations.

Background and Problem

• Problem / challenge: Media portrayals of people with disabilities (PwD) are historically absent, stereotyped, or inaccurate. AI-generated images trained on biased datasets perpetuate these issues, producing unrealistic and dehumanizing depictions of PwD.
• Significance: Accurate representation in AI-generated media can challenge stereotypes, affirm identities, and promote inclusion, impacting education, employment, and societal perceptions of PwD.
• Motivation and related work: Prior research has explored individual self-representation but lacks community-defined standards for collective representation in AI. Current AI data practices exclude communities from defining their representation, relying on biased, web-scraped datasets. This paper addresses the gap by enabling disability communities to define and curate datasets for AI.

Solution

• Proposed approach: Community Library Creator—a prototype platform that scaffolds disability communities in defining positive representation, curating datasets, and evaluating AI-generated images.
• Novelty:
  1. Technology-supported, community-led approach to defining and curating datasets for AI.
  2. Introduction of design scaffolds like Pinboard, Magazine, and Community Library for structured representation definition.
  3. Development of community-centric AI evaluation metrics based on qualitative and quantitative insights.
• Procedure and key techniques:
  • Phase 1: Define representation through Pinboard (image selection and reflection), Magazine (prioritizing themes), and community outreach.
  • Phase 2: Curate a balanced Community Library of 400 images with annotations, including bounding boxes and auto-generated prompts.
  • Phase 3: Evaluate AI-generated images using community-defined prompts and holistic rating tasks.

Results

• Concrete findings:
  • Communities curated thematic datasets with ∼400 images, reflecting diverse representation themes like education, work, and cultural heritage.
  • AI evaluation revealed nuanced criteria for assessing images, including realism, respectfulness, and alignment with prompts.
  • Project leads rated AI-generated images across a 5-point scale, capturing preferences and qualitative insights.
• Advantage over baselines:
  • Community-defined datasets and evaluation metrics offer richer, context-specific insights compared to generic or crowd-sourced annotations.
  • Focus on positive representation avoids extractive practices seen in red-teaming or harm-centric approaches.
• Experiments / evaluation:
  • Collaborations with three disability advocacy organizations representing people with dwarfism and vision impairments across the Global North and South.
  • AI evaluation included ∼300 images generated from three models (GPT-Image-1, Imagen4-Ultra, Stable Diffusion 3.5 Large Turbo).
  • Mixed methods: qualitative content analysis of transcripts, system entries, and community feedback.
• Limitations and future work:
  • Limited generalizability due to engagement with three organizations; broader studies needed.
  • Model-specific errors were not targeted; future work could address model-specific limitations.
  • Open questions on scaling datasets for pre-training and leveraging them for model alignment.

Summary

This paper introduces the Community Library Creator, a prototype platform enabling disability communities to define positive representation, curate datasets, and evaluate AI-generated images. Collaborations with three advocacy organizations produced thematic datasets and nuanced evaluation metrics, addressing historical biases in AI media generation. Findings highlight the interpretative and contextual nature of representation, emphasizing community agency in shaping AI data practices. Future work should explore broader applicability and model-specific adaptations to improve inclusivity in generative AI systems.