Your AI Can't See 16% of Humanity. Here's the 5-Minute Check That Proves It.

AI accessibility is improving fast—but it still erases disability when models are trained on incomplete data and when products treat accessibility as a checklist. Elsa Honison, a deaf-blind disability advocate and writer, describes how early image-generation tools repeatedly “corrected” her into a non-disabled appearance, while newer systems like ChatGPT and Claude increasingly respect disability traits in outputs—an advance that matters because representation affects inclusion, community belonging, and whether disabled people can safely use AI without feeling misseen.

Honison’s turning point came from experimenting with early Microsoft Copilot image features. When she asked for a selfie-style image that included her hearing aids and cataracted eye, the first outputs were distorted and unsettling; later versions gave children blindness but not adults, and the system kept apologizing for her disability. More importantly, it refused to render her “blind eye” as she experiences it, repeatedly changing it into two eyes. That pattern—disability being treated as an error to fix—became a prompt for deeper testing and advocacy.

Over the past 24 months, Honison says the tone and accuracy have shifted. ChatGPT and Claude no longer apologize for disability, and Claude can answer questions about disability language choices, including person-first versus identity-first preferences. She also cites a BBC article about a prosthetic wearer using AI to draw prosthetics; after trying again, she found AI could “mostly” render her cataracted eye and hearing aids depending on the prompt, without trying to convert her into a non-disabled person. The implication is not just better graphics, but a growing ability to treat disability as a normal visual and informational attribute rather than something to remove.

The conversation widens from image generation to the broader risk that AI systems lack disabled people in training data. Honison points to MIT’s Morality Machine (2019), which tested trolley-problem ethics using bodies crossing streets but reportedly had no disabled people in training data—meaning wheelchairs and white canes were absent. That gap forces models to reason about “disabled” outcomes using proxies like old age or children, revealing how missing representation can become dangerous.

At the same time, AI can be a powerful adaptive aid when it puts control in the hands of disabled users. Honison describes practical uses: taking pictures of distant text and having AI read it as large print; zeroing in on specific handwriting within a packet; reading prescription bottles; and replacing volunteer-dependent tools like Be My Eyes with AI-driven assistance that avoids privacy and safety concerns. For ADHD, she notes people use AI agents to track medications and plan next steps. The common thread is autonomy: AI can “fill in the spaces” without requiring constant help from another person.

For builders, Honison argues accessibility can’t be solved by WIKG-style checkboxes or “AI overlays” that assume the system will catch everything later. She warns that accessibility tech debt is real and that “fix it in post” often becomes permanent. With local inference on the horizon—glasses-like always-on layers and LLM-friendly chips—there’s a temptation for app makers to punt accessibility to the overlay, but that approach still fails when users’ needs vary and when AI can’t be forced to adapt for every disability. She also calls out limitations of screenshot-based accessibility critiques: an LLM may detect contrast but miss whether links are properly labeled or whether content is readable by screen readers.

Finally, she urges AI-assisted development tools, including vibe coding platforms like Lovable, to bake accessibility “nuts and bolts” into the workflow so new builders learn the skills rather than relying on everyone to guess what disabled users need. Her closing advice is blunt: don’t trust ChatGPT to write perfectly accessible code—double-check it—and keep disabled people visible in the systems being built.