Creating Disability Story Videos with Generative AI: Motivation, Expression, and Sharing
Shuo Niu, Dylan Clements, Hyungsin Kim · 2026 · Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI '26) · doi:10.1145/3772318.3791495
Summary
Niu, Clements, and Kim worked with nine people with disabilities (PwDs) from a Massachusetts disability advocacy group (pseudonymized as 'Campaign') to study how novice users adopt generative AI for creating first-person disability storytelling videos. Participants had disabilities spanning cerebral palsy, autism spectrum disorder, intellectual disability, ADHD, epilepsy, scoliosis, low muscle tone, and visual impairment, and most had no prior GenAI experience. Over 1.5-hour Zoom sessions each participant produced a ~1-minute video about a disability experience using a three-stage pipeline: ChatGPT (GPT-4o) generated a six-scene script, DALL-E 2 produced one image per scene, ElevenLabs generated voiceovers, and CapCut assembled the final video. The researcher operated the tools while participants dictated prompts, refined outputs, and made creative decisions — a setup designed to lower barriers for PwDs with limited media-production experience. The study is grounded in Lambert's seven-step theory of digital storytelling and investigates three questions: what motivations and moments PwDs choose to depict, how GenAI supports the co-creation of story components, and how PwDs decide whether to share AI-generated disability stories. Data consisted of anonymized ChatGPT conversation logs, session recordings, and post-study semi-structured interviews analyzed through thematic analysis with affinity diagramming, achieving substantial inter-rater agreement (Krippendorff's alpha 0.78-0.99 across coding tasks).
Key findings
Participants used GenAI to depict real-world accessibility barriers (inaccessible restaurants, Boston Green Line stations without ramps, State House hearings without access), envisioned futures, gratitude toward allies, and mental-health struggles — often framing videos with explicit advocacy call-outs ('Accessibility is not optional'; 'Accessibility is not a request, it is a right'). Most participants chose fictional character names rather than their real identities, using GenAI as a 'selective disclosure' tool that lets the experience remain authentic while the persona is fabricated. Across sessions, GenAI frequently misrepresented disability: DALL-E defaulted to wheelchairs for cerebral palsy, refused to render a posterior walker, depicted visually impaired characters in sunglasses (a stereotype), failed to place braille displays correctly, and produced inconsistent clothing, faces, and environments across scenes. Participants nonetheless valued GenAI for articulating emotions (facial expressions, body language), starting stories for writers who 'crumple up drafts,' and visualizing scenarios too risky or inaccessible to film. Sharing intentions split across three audiences: the general public on social media (for awareness), family and friends (to explain their lives), and property administrators (to prompt accessibility retrofits). The authors synthesize findings into a 'Momentous Depiction' framework with four GenAI affordances: non-capturable depiction, identity representation and non-disclosure, context realism and consistency, and emotion and social experience articulation.
Relevance
For practitioners designing AI-assisted content tools, this paper concretely shows where current GenAI fails PwD creators — inaccurate assistive-technology rendering, default stereotypes, and visual inconsistency between scenes — and offers three actionable design directions: position GenAI as a 'story completer' rather than a 'story writer' so PwDs retain narrative authorship; support multimodal media formats beyond static images plus monologue voiceover; and build 'imperfection correctors' that let creators lock character traits, clothing, and assistive technologies across scenes. The selective-disclosure finding is especially useful for social-media and advocacy-tool designers: GenAI can let creators share lived experience without exposing identifiable images. Limitations include the small, regionally homogeneous sample, focus on novice creators, and reliance on tools (GPT-4o, DALL-E 2, ElevenLabs) that predate newer video models like Sora 2 and Veo 3.1. Even so, the Momentous Depiction framework provides a useful lens for evaluating how future GenAI handles disability representation.
Tags: generative AI · disability storytelling · video accessibility · disability advocacy · LLM · content creation · disability identity · AI bias · HCI · qualitative research