Some claim (their) AI can fix accessibility automatically and as much as some want to believe this to be an absolute truth, we should be realistic and write about it more.
I accept the fact that a lot of AI is not sustainable, is actively stealing content from us to sell it back to us. But it seems that we came too far to deny the positive sides as well. Personally, I hate the idea of this grand theft of human knowledge, especially of content that is not in public domain. But the technology powering this unethical machine can still be used for good, so I also see that we can turn it into our advantage with local, private AI, that has our context and is empowering us and not some third parties.
Compliance is a lot about details
Now first we need to consider what the promise of accessibility compliance actually bears before we can even reflect on the AI possibilities for it. It means a website or mobile app (or document etc.) needs to comply with accessibility regulations like for example European Accessibility Act (EAA). This is the first issue, EAA is only one of many possibilities. Different countries can have different accessibility compliance requirements. Sure, EAA tries to harmonize requirements in the EU for the scope of it, but in reality we still have local differences in some countries. So claiming to have accessibility compliance can actually mean a bit different things, especially depending on the local requirements. The details are vital when it comes to compliance.
In Norway – an online shop still has to comply to legislation that requires WCAG 2.0 on levels A and AA, for example (EAA is still not incorporated). If same shop actually is compliant in Norway (I still have to find one that is, even to WCAG 2.0) it most probably is not compliant in Sweden (as their accessibility compliance is based on European Accessibility Act and stretches beyond WCAG with EN 301 549 requirements).
So same shop can, in theory, be accessibility compliant in Norway but not in Sweden. But perhaps it is compliant in Finland (as sometimes local standard interpretations may be different).
We can say that taking the strictest interpretation would cover all of the less stricter ones, so AI should just take WCAG 2.2 on levels A and AA and the rest of EN 301 549 and fix it.
In theory, that could work. In practice, AI can’t really do it automatically. Only people that understand the accessibility implementation details and most importantly how to actually verify accessibility manually could potentially be more efficient with AI to “fix compliance”. Trusting AI to just “fix accessibility” will fail without exception.
Compliance often requires more than just conformance
Compliance is often way more than just conformance to some technical standards. It requires the organization to have an accessibility program that is followed up, not just a single project that will “fix accessibility compliance”.
Can external AI solution just “fix” organizational issues? Even if it has the potential in theory, we can not just trust it. AI can for sure help people be more efficient when people bear the knowledge and responsibility. AI can augment and improve things that people understand, faster.
A lot of organizations struggle with decentralized parts that live their own lives, on both technical and organizational levels. It takes an organizational change management to fix the barriers. It takes time and training and measurements. Sure, AI can help with all of them, but it is first and foremost people that need to lead people. Augmented? Sure, if it works for them.
Conformance to technical standards is therefore not enough for compliance. At least not when compliance requirements are beyond pure technical ones. And in most cases they are not focusing only on the technical parts. Even if we only have to have an updated accessibility statement – it does not help to just programmatically update the updated at date dynamically. If we let AI to “audit” and report findings to the statement without checking the results, people will figure it out and there goes our trust. It may appear that our statement is updated, but people with disabilities will soon figure out the bluff. As will the authorities.
Realization of plans is not just to generate plan text with AI. “You are an accessibility expert, audit this for EN 301 549” will produce wrong info. Accessibility needs real, human, effort. Even when AI agents are the only ones contributing to design and code and people are “just checking” the new code and changes, we still absolutely need knowledge and understanding.
AI can help specialists to be more effective. Sometimes it can even help to teach (but we can get wrong info, so please be careful). But ultimately we need people to check if final product is really accessible. Until AI will really have all the best practices, access to assistive technologies directly, real context and vetted opinions we can’t just leave it alone.
If you don’t have the knowledge to verify, do not trust AI
I guess this is quite obvious to anyone testing out AI with subjects they really understand well. When majority of AI produced text/code/image/video is quite satisfactory, but something feels off and when we really dig into it we figure out a fatal detail. A detail easily missed by people not familiar with the subject, but a detail that could break the whole thing.
When you know, understand – you can fix it. And you have hopefully saved time (on expense of tokens). If you do not know – and just “approve” it as it seems legit – you have potentially “poisoned the well” that somebody else will drink from. This example stretches beyond accessibility, but I have seen lots of proof of it, even with shamelessly priced models that outrank any specialists time rate.
And this is really scary. Trusting best in class AI that is more expensive than veteran specialists for same task and still getting details wrong should be forbidden by law. But it’s impossible to enforce such a law. Especially when you agreed with the terms of service that claims no responsibility.
Nevertheless – AI can be used for good if people operating it have enough knowledge. So out-sourcing knowledge is the real danger here. And learning just with AI is dangerous, as we may learn the same problematic details and be the agents of AI instead the other way around.
AI for accessibility compliance – can help when people have the knowledge, domain understanding, can and actually do verify manually. AI for accessibility compliance without people that can actually verify – a disaster waiting to happen. Often more expensive than real people. Maybe not at start, but often after some time, less evidently but more in the shadows.
