- From: Léonie Watson via GitHub <noreply@w3.org>
- Date: Wed, 11 Mar 2026 11:21:43 +0000
- To: public-webai@w3.org
A few initial thoughts... The data most models are trained on is a risk. Models trained on [common crawl]() data have been taught using a web that still [struggles with accessibility]() and consequently, the accessibility mistakes now being made by chatbots and agents bear a striking resemblance to the mistakes made by designers and developers. There may be an opportunity to counter this risk, but it needs more exploration. Given better training data, robust skills, clear guardrails etc., might it be possible for AI agents to assist designers and developers more reliably in the creation of accessible interfaces? Another risk with the same cause is the representation of People With Disabilities (PWD). The training data includes all the misinformation, misconceptions, and mischaracterisations of PWD that has accumulated over time. For example, received wisdom has it that most blind people read Braille and/or have a service animal, when in fact both of those groups are small minorities within the blind/low vision community. An opportunity that is already evident is the use of computer vision/image recognition in AI models. Browsers have had the ability to request a description of images without alternative textsfor some time, and more recently screen readers have introduced the ability for users to query popular AI models for text descriptions (with the added benefit of being able to ask follow-up questions to derive more specific or detailed information). Another evident opportunity is the ability to access content from within large, complex, or inaccessible documents. Until the advent of recent AI models, the only alternative to an inaccessible PDF for a screen reader user was to use Optical Character Recognition (OCR) to extract the text content. This is a useful solution but it's a "blunt instrument". Now it's possible to upload the document to a chatbot and prompt it to extract the information you need in the form or format you need it. Similarly, the ability to use AI models to summarise complex documents can be helpful to neurodiverse people who find it difficult to consume or process large amounts of content or interpret abstract concepts. For people who find reading difficult, the conversational capabilities available in many chatbots can lower the barriers. The risk of hallucinations is, of course, ever present, but that is not in and of itself an accessibility risk. -- GitHub Notification of comment by LJWatson Please view or discuss this issue at https://github.com/w3c/webai/issues/14#issuecomment-4038521882 using your GitHub account -- Sent via github-notify-ml as configured in https://github.com/w3c/github-notify-ml-config
Received on Wednesday, 11 March 2026 11:21:44 UTC