Key Takeaways
- • Disabled people appear in only 2.3% of stock photo training data
- • AI models fail to render realistic prosthetics in 78% of attempts
- • Wheelchairs are anatomically impossible in 65% of AI generations
- • Curated inclusive datasets improve representation accuracy by 340%
- • Community-driven fine-tuning shows promising results in early pilots
The representation gap in AI imagery
AI image generation models trained on internet-scraped data inherit the biases present in their training sets. Research by AI Now Institute shows disabled people appear in only 2.3% of commonly used training datasets. This manifests in poor representation, often producing stereotyped, medicalized, or completely absent depictions.
Training data limitations
Stock photo databases and social media—primary sources for AI training—systematically underrepresent disabled people. When disabilities do appear, they often follow narrow tropes: wheelchair users in medical contexts, inspiration porn narratives, or tragic framings.
Model failures in disability representation
- Prosthetics: AI models frequently render unrealistic or anatomically impossible prosthetic devices.
- Mobility aids: Wheelchairs, crutches, and walkers are often distorted or physically implausible.
- Visible differences: Scars, limb differences, and other visible disabilities are smoothed away or rendered inconsistently.
Impact on disabled communities
Poor AI representation perpetuates exclusion in media, marketing, and creative industries. As AI-generated imagery becomes more prevalent, the absence of authentic disability representation compounds existing marginalization.
Paths toward improvement
Researchers and advocates are working toward better representation:
- Curated datasets featuring authentic disability representation
- Collaboration with disabled artists and photographers in training data creation
- Explicit bias testing and auditing of model outputs
- Community-driven fine-tuning of models for specific disability contexts
Ethical considerations
Improving AI disability representation must center disabled voices and avoid perpetuating new harms through well-intentioned but misguided efforts. Nothing about us without us applies equally to AI development.
Learn more about AI bias in our AI ethics section and explore our AI tools hub.
