There is an expression that is used more and more when referencing use of AI, and makes me uncomfortable: “Keeping the human in the loop”. I, instead, believe we need to ensure that we “Keep the human in the middle”. What is more relevant with this then, other than reflecting on the impact, negative but positive too, of AI against digital accessibility?

Is AI improving or impeding web accessibility?

Sadly, the initial look is mostly negative: A lot of AI-generated code and content is inaccessible by default.

The team behind US-based WebAccessBench have been benchmarking AI-generated UI outputs, and the results are not great. Their findings describe what they call a “systematic civil-rights failure”: AI models are trained on over 20 years of internet content, and that content has historically been riddled with accessibility barriers. So when AI learns from it and then generates new interfaces and content, it faithfully reproduces those same barriers. This isn’t a minor bug that can just be fixed, it’s embedded in the structure and nature of how AI is trained, thinks and responds.

AI is locking people out. At Scale. – Digital Accessibility Reliability in LLM-Generated Web Interfaces, including white paper, by Casey Kreer.

As it is the reality with all things related to AI, everything is significantly faster and is applied at a greater scale. Inaccessible patterns are spreading faster than they can be reviewed and fixed. This is not a risk of failure complying with related regulations and legislation, but a failure in inclusion. Specifically for our sector, this could result in audiences unable to access our University website and admission pages to understand our study and research offerings, our student portal to find the information that they need, or our help and support pages to identify who to contact to get the support they need.

In the University of Edinburgh we have issued relevant guidance to support our staff & students make the right choices in using AI.

Guidance for student and staff on AI use

Looking in the wider sector, though, and more globally, the risk is real. The World Wide Web is accessed by many people requiring the content, and interfaces, to be accessible, so all of us need to keep a keen eye to ensure that. That has become even more important in the age of AI.

W3C is reviewing the impact of AI against its standards

Thankfully, the broader web standards community is taking this seriously. The W3C’s Accessible Platform Architectures Working Group has published an editor’s draft on the accessibility of machine learning and generative AI, which is attempting to keep the discussion alive and, hopefully, set the standard of expectations.

Accessibility of machine learning and generative AI – Editor’s draft, W3C

The document is honest about where AI currently falls short. The automation of alt text, one of the key accessibility features, is getting better, but is still inconsistent. For example, a bar chart showing children’s favourite colours might be described as simply “a graph with different coloured bars” by one tool, while a more sophisticated model provides a rich, detailed breakdown of the data, as it is appropriate. The gap between those two outputs is important, since it highlights the difference between understanding and exclusion for someone relying on a screen reader.

Automatic speech recognition for captions is another area where progress is real but the finish line is still some way off. A person experiencing a hearing deficiency will not consider an 85% accuracy rate good enough when captions are the primary means of accessing content.

Additionally, the European Union AI Act has been published since the 1st of August 2024. It makes direct accessibility-related references in terms of scope, requirement for content to be accessible and consideration of vulnerable groups in application of AI. It still, though, relies on existing EU law, like the European Accessibility Act to set the expectations, which makes sense since no one wants to reinvent the wheel.

EU AI Act – First regulation of Artificial Intelligence

AI as a positive agent for accessibility

Moving away from the negative aspects, it is true that AI holds genuine promise to support better digital accessibility. Real-time remediation of colour contrast issues, intelligent identification of non-descriptive links, improved heading structure detection, plain language simplification are all areas where machine learning and generative AI could make a tangible difference, in real time, particularly for users with cognitive disabilities or those who rely on assistive technologies.

Joshua Morast, a web platforms and accessibility specialist, published a fascinating experiment: he asked AI to simulate what using a screen reader on a real application feels like. Not to replace testing, but to generate a detailed transcript of the screen reader experience that he could share with stakeholders who had never used one.

I asked AI to use a Screen Reader – Blog post by Joshua Morast

The results were fascinating, to say the least. The AI produced a long, detailed, and directionally accurate simulation of the experience of user navigating through a government permit application using a screen reader. It identified several issues: skip links that went nowhere, navigation drop down menus whose triggers were entirely unreachable by keyboard, and link text repeated identically ten times with no distinguishing context. The summary came back with the note that “a screen reader user attempting to search for a building permit would need to tab past approximately 40 interactive elements”. Not only a failure of accessibility, but of user experience altogether.

Joshua is clear that this approach doesn’t replace human testing or the specialist expertise of someone who actually uses assistive technology day to day.

Apart from accelerating the understanding of content challenges for screen reader users, there is another, hidden, benefit: It is often very difficult to communicate the challenge of abstract accessibility principles. “WCAG 2.1 AA compliance” doesn’t mean much to someone who has never relied on a screen reader, or is an accessibility expert. Even though there are applications that simulate the experience, they require time to set up and use. A simulated transcript of what their service actually announces, written out in plain text, is something else entirely. It can very quickly reveal fundamental issues which otherwise might not have triggered the necessary action.

There is a yin and there is a yang

AI is neither the villain nor the saviour of web accessibility. It’s a tool, and like most tools, its impact depends entirely on how it’s used and who’s in charge of it.

From one side, it carries over the biases and deficiencies of a 20+ year old World Wide Web. On the other hand, experiments like the one by Joshua Morast, show that AI can become a powerful assistive layer to accessibility experts. It cannot replace the human expertise in accessibility. Furthermore, it cannot be a replacement for continue having an approach of accessibility by design.

Personally, I will continue to seek how we can use AI as an ally in our efforts to deliver a more accessible University of Edinburgh web estate, and post my thoughts and findings.

I would love to hear how others in the Higher Education community, or beyond, are thinking about this. Are you already using AI tools in your accessibility workflows? What’s working? What’s causing concern? Please, drop a comment below or get in touch.