AI is here. Friend or Foe? (Part 6 of 7)
Accessibility TestingVocation
AI is here.
In a recent survey I put up on LinkedIn, asking whether Labrador should have an AI remediation tool, a third of people said they'd hate it, another third said they'd love it and the remaining third said "meh." Talk about split opinion!
When I was starting in accessibility I think I would have loved some of the benefits of AI. Everyone else around me seemed to know more than me and I was never confident reading code let alone suggesting fixes for it. This is something that AI is pretty good at to be fair. I could have learned a lot and quickly. Some will point out that AI still makes mistakes, and it definitely does. But those mistakes are increasingly rare, and despite that it can still be a great tool to learn how to turn bad code into good code and to put WCAG requirements into plain language.
So how do I feel about it now, today, I use AI every day. I use it in my audit work, I use it when I'm writing, I use it when I'm researching, I use it when I'm trying to work out what I think about something. Oddly though, despite this, I'm not really a fan of it. I can feel it rotting my brain in the same way that social media does and that's something I'm trying to get a handle on. I also hate how data centres are springing up every five minutes and the impact they have on the communities that live right next door.
But I have to accept that it's here and it's not going away and my work has changed. and just like the introduction of the abacus, the calculator, the desktop computer and mobile phones it's use is unavoidable now, but if it stopped working tomorrow I'd be fine. I managed for 20+ years in this industry without it.
The context you're walking into
Before I get into any of it, a bit of context on why the noise around AI in accessibility is so loud right now, besides the fact that it's new and shiny.
The European Accessibility Act came into force in 2025, which means a very large market of organisations who previously weren't thinking about accessibility were suddenly required to. That's a huge commercial opportunity, and where there's a huge commercial opportunity there's a corresponding rush of vendors making increasingly fantastic claims about what their AI can do.
One esteemed colleague in the industry made the claim early last year that 100% of accessibility testing could be done by their AI before the end of the year (claim made in March 2025), and that 57% of it could already be done by their tools today. This isn't the first claim like this and it won't be the last. It reads a bit like the overlay-widget claims from a few years ago and it went down about as well. Thanks you guys. We're still waiting.
I mention this because if you're new to the field, you're going to hear a lot of these claims. From vendors. From LinkedIn thought leaders. From clients repeating what they heard from vendors. Don't panic, it's all BS.
Last year I offered a challenge that nobody has taken up: Show me a single WCAG 2.2 AA criterion that AI can fully, flawlessly, reliably test today. Not partially. Not with a human checking the output. Fully. Just one.
In reality I think the day may come, but not anytime soon.
What AI is actually good for
The bits of my work where AI has made a real difference are the bits that used to slow me down without adding much value.
Remediation advice. If I find a form field with no label, I know how to fix it. But if the fix involves a nested component, a legacy framework, and some unusual ARIA relationships, I could spend a long time working out the right recommendation. Now I can describe the situation to a model, get a starting point, sense-check it against what I know, and move on. The time saving over a full audit scope is significant. This for me is the main place where AI is useful in accessibility testing.
Explaining things in plain language. I sometimes struggle to translate a technical finding into something a stakeholder can act on. AI is decent at that first pass. I still edit it heavily, because the tone is usually wrong, but the raw material is useful.
Research. Whether or not you're using AI, you'll still need to do research, this might just be about the types of AI features that are supported on Mac-Safari vs Mac-Chrome. I find that AI can be useful in finding and summarising real-world, verifiable articles that contain this type of information.
Getting unstuck in writing. When I'm halfway through writing something and I've hit a wall. I'll paste my draft in and ask for suggestions. I often don't use any of them. But seeing them helps me work out what I want or don't want to say.
The pattern in all of these is the same. AI is useful for the friction bits. The bits where I know what I want but can't quite get there. It's not useful for the bits that are actually the work.
What it isn't good for
Manual accessibility testing.
AI cannot do what a human tester does. Not now, and I'd argue not for a very long time. There are things it can't perceive the way a person does. There are things it can't judge because judgment isn't pattern matching. It's a Large Language Model with no actual experience of using assistive technology, or experiencing using anything like a human does and it never gets annoyed or frustrated.
An AI can tell you whether an image has alt text. It cannot tell you whether that alt text is any good. It can try, but it still doesn't always do a good job. It can tell you a button has an accessible name. It cannot tell you whether the name makes sense in the context of the page you're on.
Automation is very good at the first mile of a process. It's spectacularly bad at the last mile. That's why factories still have humans on assembly lines even after decades of automation. Accessibility is no different. Automated tools, AI-powered or otherwise, handle the first mile of finding obvious issues. The last mile, working out whether real people can actually use the thing, is a human problem.
The gap between "compliant on paper" and "actually usable" is where accessibility really lives. And that gap is a human gap. It's about attention, context, empathy, and the feeling that something isn't right even though you can't immediately say why. AI doesn't have that feeling. It can only pretend.
"Will AI take my new accessibility career before it starts?"
I really don't think so.
Longer answer: the parts of accessibility work that AI can do are the parts that were already the least valuable and least enjoyable.
Automated scanning has been around for decades. A bit of trivia, The first one I used was called 'Bobby' which was launched in 1996. Today, modern tools catch 30% of issues, mostly the boring ones. AI is making automated scanning a bit better and a bit faster, and that's welcome. But automated scanning has never been the actual job. Most AI tools I've looked at build their knowledge on algorithmic tools that are already in the market; meaning that they are also limited to that 30% ceiling which seems very hard to break through.
The actual job is manual testing. Watching real behaviour. Educating yourself on how users really use your digital products. Making judgment calls. Writing something a developer can act on and helping them understand why it works that way. Delivering a report playback. Sitting in a delivery meeting and helping a nervous product manager understand why the thing they built isn't going to work for a chunk of their users. As a consultant a lot of what I do is steering many different people with many different priorities and objectives in the same direction. That's a very human thing.
In summary, when used appropriately, AI raises the value of good human testers, because it reduces friction for the parts of the job that are tedious. So, ignoring how I feel about the negative impacts of AI. If you're entering this field today, you couldn't have picked a better time.
"Will AI-generated audits flood the market and lower standards?"
Some people will use AI to produce cheap, bad audits and sell them to clients who don't know the difference. That's already happening. They're worse than useless. They are dangerous. They're often confidently wrong in ways that make the client feel like they've done their bit when they really haven't. This has been happening with automated testing tools and cheap-and-nasty overlays or a long time. it's not new. It's a strategy employed by big business who want to sell quick fixes. But they make sure their contracts say they aren't liable when you end up in court.
Automated-scan-dressed-up-as-a-report has been a thing for years. Accessibility overlays (web accessibility widgets you find on websites) have been disrupting the market with sub-par "fixes" that have landed their clients in court and facing very expensive class-action lawsuits. What AI does is make the bad reports look a bit more polished, which arguably makes them even more dangerous because the mistakes are harder to spot.
Good work is visible over time. Clients who get burnt by a bad audit come looking for someone better next time. Word travels. The market for accessibility work is small enough that reputation matters more than marketing.
Your job, if you're worried about this, is to be one of the ones doing it properly. Do audits that stand up to hard scrutiny. Deliver them well. Follow up. Build the kind of reputation that survives the noise. That's it. The AI-slop merchants will do their thing, and they will lose the clients who care, and those clients will come to people like you.
"Am I falling behind if I'm not using AI heavily?"
This is just my opinion. I might be wrong. I think that If you're not using AI at all in your work in 2026, you're going to feel behind in about a year. Not because AI has replaced anything important, but because the productivity gap between people using it well and people not using it at all is real, and it's growing. If a competitor of yours can turn round a scoped audit in three days and you take five, over time that will show up in bids. This isn't just applicable to accessibility, it's applicable to most jobs. I'm not promoting AI. I've said how I feel about it, but this is the reality, at least the way that I'm perceiving it.
But heavy use of AI, for its own sake, is not the goal. In fact if you go down this route the quality of your work is likely to be absolutely dreadful. The goal is to know what it can do, know what it can't, and use it for the bits where it saves you time without making you worse at your job.
If you're brand new to accessibility don't lead with AI. Learn the craft first. Do your first audits by hand. Feel the shape of the work. Notice where the friction is. Then, when you know what you're doing, introduce AI to smooth out the specific bits where you get stuck. Maybe it's the remediation bit. But pause to understand the answers that AI gives you and don't just paste them in without checking. If you start with AI, you'll never build the underlying judgment that tells you when it's got something wrong. And it will get things wrong. And it will tell you with complete confidence that it's answers are perfect, and reliable and smell like roses and that your decisions are always the best ones.
This is the specific danger for early career testers. AI is very good at sounding like it knows what it's talking about. If you don't have your own knowledge to check it against, you will trust it, and without any doubt in my mind, it will let you down.
What I actually do
Here's roughly what my workflow looks like for a typical audit
I don't use AI to find accessibility issues. I have tried. Accessibility testing can be incredibly tedious and I have put a lot of time and effort into trying third-party tools and developing my own tools to test for accessibility issues. It would be lovely to reduce the workload to something like a light QA check. It just isn't possible right now.
I can say with a high degree of confidence that you will spend more time correcting the AI assessment than you would finding the issues yourself. It will burn through tokens like matchsticks to make up some unbelievable nonsense that you'll have to bin, or that you would have spotted in seconds with 100% confidence.
I use my eyes, a keyboard, a screen reader, some tried-and-tested automated testing tools which still make mistakes, but I've learned to know what to trust. And of course I use Labrador to structure the process.
When I've found an issue and I'm working out the fix, I'll sometimes ask AI to look at it. I'll feed it the specific code and ask what's up. I know it will check it against accessible design patterns that have been written by humans (and stolen by AI - another issue we won't get into here). I read it. I sense-check it. It saves me minutes per finding, and over an audit that adds up.
When I'm researching something, especially something I'm not fully sure about, I'll normally start with AI to get oriented.
That's it. It's not glamorous. It's a set of small, boring efficiencies. But it adds up to a workflow that's meaningfully faster than what I did five years ago.
A thought on where this should go
The AI tools being built for accessibility right now are mostly being built by individual vendors in isolation. Each one training on its own dataset, keeping its models secret and proprietary, competing to be the one that dominates the market. I understand the commercial logic of that. But it's not the outcome that's best for disabled users, and I don't think it's the outcome that's best for the profession either.
A more ethical version of this where a community of accessibility professionals, developers, and disabled users shared anonymised data about the issues we find and how we fix them. A shared training dataset. Better, richer, more diverse than anything one company could build alone. Consultants and users helping label and verify examples. Common tagging so models built by different people can compare notes.
Richer training data equals better AI. Community validation catches the mistakes any single vendor's model would miss. Shared standards mean we don't end up with ten incompatible flavours of the same tool.
I'm thinking out loud here, and I don't have a plan for how this could happen. But if you're at the start of your career and you're thinking about where to put your energy, this is one of the more interesting questions in the field. The people who work out how to do accessibility AI collaboratively rather than competitively are going to shape what this profession looks like in five years time. That's what I think anyway.
The uncomfortable bit
I'm building a product (Labrador) that uses AI in various ways. I use AI myself, daily. I think it's a genuinely useful tool in the hands of a skilled professional. But I also worry, about what it's doing to the profession as a whole. Society even.
What I will say is this; The way you handle AI in your career is a values choice, not just a productivity choice. Decide what you want your work to feel like. Decide what your standard is. Then use AI, or don't, in the service of that.
Where next
Part 7 is the last one in this series. The five year roadmap. What the arc actually looks like from the inside. What you'll be doing at year one, year three, year five, and how the work changes when you stop feeling like you're catching up and start feeling like you know the terrain.
See you there.