AI impacts on Accessibility
Mixed bag of good and bad
I recently had the honour and pleasure of hosting Neil Jarvis on a joint webinar sponsored by Squiz, discussing digital accessibility. Neil, besides being an expert in the field, also has a lived experience — he’s completely blind, and relies on assistive technology.
We were particularly interested in the impacts AI has on the experience of people with accessibility issues across digital systems, but the discussion went wider than that. Below are some of my thoughts of the subject, in the hope that it helps raise awareness for this important subject.
Digital Accessibility
Both Neil and I have been around to see the rise of the PC and Internet, mobile phones, and now AI, so this isn’t the first technology revolution — or hype cycle — we’ve seen play out. While the pace is increasing, it’s interesting to compare how each wave has transformed the experiences of people with disabilities and accessibility issues.
I’ve asked Neil about the good and the bad in AI. Much of his interactions with AI are via a mobile device, and he is using dedicated apps as well as ones from all the major AI providers. Some of the wonderful developments is being able to point a phone’s camera and get a verbal description of a scene. He recounted how he can now “look” at pictures back from his childhood or his wedding, and have the verbal descriptions revive and relive those pleasant memories. It is something that we take for granted, so take a moment to appreciate being able to ‘see’ mementos this way. Other apps help with daily lives, like checking whether an outfit ‘goes together’ and isn’t a fashion disaster.
Of course, AI is prone to hallucinations. While I could certainly use that fashion advice myself (assuming the AI may be correct more often), I recall when we had a friend who was temporarily blind. She used a similar app which, when describing my wife, decided to very authoritatively add two years to her ago. Hilarity ensued. It can also be more problematic, as when it described to Neil an object being on the left, when in fact it was on the right. (And in a typical fashion, pointing the error to the AI made it argue back like a toddler). Luckily it was in an environment he was familiar with, and could catch the error without harm.
These kind of applications are a good case where even having something that isn’t super-accurate is an improvement. One still needs to verify the output, but it’s a marked improvement in the quality of life.
Digital Divide
Neil is technically savvy, understands how to use the tools, and can work with them. That is not always the case. Digital literacy, financial resources, and other similar aspects beyond physical disabilities, can greatly affect people’s access and derived value from technology. For those vulnerable populations, increased reliance on technology deepens the divide between those who can and those who struggle.
Just think about trying to contact a utility’s call centre, and being constantly redirected to do your business on the web. It’s not helpful when the website isn’t accessible, or when an automated agent gives you wrong advice and service. Or think of the subscription fees all those apps and services require. The people who need them the most, those with disabilities, are often in lower income brackets due to struggles with full time employment. Those new apps and associated fees may keep them out of reach of the people who would benefit the most, which again exacerbates the situation.
Automated Accessibility
So you want to build an accessible product or service. How do you do it?
A big problem with many automated accessibility checkers is that they only catch the 30-40% of errors that can be detected automatically. False positives and negatives also abound, for example an image with the alt-text of “alt text goes here” would pass the check, while other issues may be triggered which are fine. There is no replacement for having a human who understands the experience review it.
Adding AI into the mix has its own sets of issues. Asking an AI coding agent to build a site that is accessible is like asking it to build a site that is secure; it’ll happily insist it did so, regardless of the actual code. And AI-based interfaces themselves have issues, from the direct accessibility to screen-readers to the content itself that may be harder to verify and increase the digital divide.
That said, there are certainly cases where AI technologies can be used to enhance accessibility. This is a topic we at Squiz are passionate about, as part of our overall mission of improving digital experiences to enhance off-line lives. AI tools can find unfit alt-text and help correct it. They can help authors “translate” the text written by experts and approved by lawyers into something most humans can understand.
The trick, as always, is to understand that AI is just another tool. Just like the automated site-scanners aren’t sufficient to ensure accessibility, neither is AI a magic solution. You need to understand the capabilities of the technology (the jagged edge), it’s capacity to solve a particular issue, and keep a human expert in the loop to direct it.
The Future
Both Neil and I seem to have misplaced our crystal balls 😜
That said, even though these are times of great change and a dawn of a new type of general-purpose technology, there are some observations we can make from history that will apply to AI as well:
Humans tend to over-estimate the short-term effects and under-estimate the long term effects.
The AI hype cycle will crash at some point, due to capability and adoption gaps. The long-term effects will likely be something surprising, not what we may predict now. Social media, for example, didn’t mean the rise of world-wide connected democracies, and we’re only now realising the psychological harms of automated content.
Magic hammers and silver bullets are the stuff of fairy-tales, not reality.
AI is just another tool, which needs to be wielded by experts in their field who know what they’re doing, directing and verifying the output.
Accessibility isn’t a policy in SharePoint; it’s the lived experience, and the difference between good and bad has huge impacts on whole communities.
You can — and should — absolutely start today, and build it into your workflow. Accessibility is not a perfect state, but constantly evolving thing you need to work on
If you’d like to watch the full webinar or read a more detailed report of key insights, you can do so here: https://www.squiz.net/blog/what-organizations-can-learn-from-real-experiences-with-ai-and-accessibility
That’s it for now. This is another view on Responsible AI, which I hope you find useful. It is something I always cared about, and why I love working for a company that built it into its ethos. It goes to show that while there is plenty of bad associated with AI (suggested reading: I Love Generative AI and Hate the Companies Building It), it can be used for good — it just takes some conscious effort.
Reach out if you’d like to talk about Responsible AI, about digital accessibility, or about building digital experiences that make a positive impact.