Mirror, mirror on the wall, who is the fairest AI (for accessibility) of them all

Without a doubt – artificial intelligence (AI) – seems to be changing our lives (for better and for worse, as it’s often the case with technology). And AI will, without a doubt change accessibility, also for better and for worse, it seems. It seems that it’s all around us, added to more and more software products and services and in some cases causing big changes – from people dropping the usual websites and just using a chat interface to products that can automate interactions with computer operating system, browser and mobile applications.

Let’s stop for a moment and check some of the current predictions:

• Some claim that AI will automatically fix all accessibility issues (they claim this for years and some of them also got financial fines because of these false claims).
• Others claim that accessibility is “dead” as we will all just consume tailored interfaces and/or automate all interactions with computers so that they will just work instead of us.
• Some now claim they will fully automate accessibility testing.

Different aspects with a common denominator – AI – but using different approaches.

I’ve been using AI for coding some time now, and I appreciate off-loading all the grunt tasks, sometimes even use it as a sparing partner, but the more I use it – the more I see that it can’t be fully trusted. Now – you can say that this is just a temporary thing and it will get better and better – and I want to believe you. And I see progress, sometimes, but then next prompt destroys it again. And so on and on.

When I manually filter out the noise it sometimes produces it helps me code faster. But – even when I explicitly specify all accessibility requirements and even provide verified documentation and examples from sources I trust, even when I use custom models filled with those valuable and trustworthy sources, I still see mistakes and issues and that makes me wonder. Sure, organizations with lots of manually verified data have better chances, but I still can’t trust a tool that is sometimes good and sometimes bad – with same input.

With 95,9% of webpages out there having automatically detectable WCAG issues, with all the open source packages claiming they conform to WCAG but actually don’t even state what was tested, with overlays blocking testing tools and claiming conformance and with built in biases – I can only say that I need to remain skeptical of all those claims of AI hype. It’s just not healthy to lie to myself, but I want to remain an optimist and wait to see.

Another consideration that is often left out of this kind of debates is how these claims may also bring a lot of bad things. People before me already reflected on them – built in biases will hurt even more people, funding for human resources will be redirected to products or minimized as “we will just let AI do it”, some organizations will most probably just buy a product instead of integrating inclusion and accessibility into their culture and we risk to make the accessibility gap even larger if people will need to buy another subscription to make their experience accessible when it should already be accessible in the first place.

I can’t predict the future, but it seems that others think they can – can’t claim that I know their reasons, but we need to ask them that they don’t try to use hypothetical solutions for real problems that discriminate people.

Even if AI consumes all written and recorded knowledge it will still miss on millions of years of evolution that got us here and is a part of us as a race, not to even mention our diversity as an individuals.

AI race is amazing, without doubt, a lot can be automated, but I am certain that we will continue to need human supervision or else risk problems. And inaccessibility can sometimes mean a wrong button label and sometimes a barrier that shuts people out completely. So – I still trust myself more than any AI – or at least I know when to ask for help which is often not the case with overconfident predictive pretrained models that “know everything and beyond”.