Could we get to the point that we can simply ask an Al tool to create an accessible website from scratch?
Not anytime soon, but eventually it could do a decent job especially as a co-author. 96.3% of the web has WCAG errors, when huge data sets are trained on inaccessible code, the code that is most likely going to reproduced is going to be inaccessible. These large language models are scraping the web, they’re being trained on poor data. They have no real anchor of what is accessible and what is not.
Without detailed testing it’s hard to be sure, however at time of writing, some casual experiments with ChatGPT on accessible UI patterns gave OK results. Asking it for a complex patterns like an accessible date picker resulted in it not knowing what to generate.
What would be more exciting is if a large model is trained exclusively on accessible code. This would be an interesting space to work in.
Will it ever be able to understand the nuances of all the possible HTML or ARIA attributes for accessibility?
Yes, but we already have lots of automated tooling that can pick up huge volumes of HTML, ARIA and accessibility errors. Unfortunately, the problem isn’t technology - it’s politics.
The politics in companies that allocates priority, allocates time, allocates resource. AI is not going to be able to alter the mindset of a CEO that chooses not to care about accessibility and disabled people.
Threads from Meta launched with many accessibility errors. This is a company with billions of pounds of resources, all of the tooling in the world at their disposal. Meta even has their own large language model. However the choice was made that accessibility and disabled people are not important. AI is not going to be able to change the mindset of certain people in power.
The CEO of Twitter fired the entire accessibility team even though he’s part of the disability community himself. Which prompts a wider thought about how to dismantle internalised ableism.
A model might be helpful to help understand the nuances of HTML and ARIA but a lot (if not all) of this information is already available. Also, it might be harder to spot errors in the answers, spreading misinformation. If teams and companies do not care to make accessibility a priority - it won’t be. AI won’t be able to help with this aspect. In fact, it might make everything much worse.
We’re already seeing the rise of overlay companies. Many of these overlay companies promise quick fixes to make code accessible, many of them also mention AI. Then, when these companies are criticised by a globally-recognised accessibility expert, they take legal action.
As with many world issues, technology can’t resolve structural and systematic problems. These have to be solved with culture change, policy, regulations, law and enforcement.
Also, use of these models have to be held legally accountable in a meaningful way - with a human being responsible. Perhaps with an AI officer, the same way companies have a dedicated data protection officer.
Could we get to the point where we can generate really impactful and meaningful alt text descriptions for images?
Yes, but with lots of damaging bias until this is taken a lot more seriously.
Amazing computer vision models are already being used on mobile and apps like and Be My Eyes, to break down these barriers.
I have mixed feelings about AI generating alt text. In some ways this could be incredibly helpful as a starting point, but only if it’s being used as a tool not to replace humans. The concern is that people, and companies, will believe AI will do all of the work and no human intervention is needed. I would worry this would promote bad practice.
We’ve seen similar issues with auto-generated captions, while they can be very useful as a starting point they still need human intervention and clean up. Captions for a fictional TV show with errors might have limited harm but this is not the same for information related to news, finance or law.
This tooling can empower disabled users to extract information from inaccessible content. Generate multiple different versions of the alt text, maybe even from different language models to get a much better idea of what the image contains.
What about the tricky nut of allowing screen reader users to explore all the data in graphs and charts and maps and infographics?
Eventually yes, however at the moment this isn’t done that well with humans. (Not because it’s not possible but due to allocating resource). Complex data, infographics and graphs need a fair amount of thought and work to make it accessible. Descriptive text is one of the best methods but data can be interpreted in many different ways.