AI1y (AI and A11y) – Part III: Can I Trust AI to Be Right 100% of the Time for Accessibility?
- Nir Horesh
- 2 minutes ago
- 4 min read
The short answer is no. The better question is: does it matter?
In the two previous posts in the AI and accessibility series, I covered:
This time I want to discuss the reliability of AI-based accessibility audits. In all the recent conversations about AI and digital accessibility, one question keeps coming up: there are no AI-based tools mature enough to be trusted 100%. I want to address this point.
AI Makes Mistakes. So Do Humans.
Let's get this out of the way: AI isn't perfect at accessibility testing. It might give the wrong alt text to an image, miss that a div should actually be a button, or add heading tags where they don't belong.
To be clear, I'm not aware of any serious tool today that even claims to perform a full accessibility audit based on AI. I do know about good tools with partial WCAG coverage, and unfortunately some bad tools that claim full coverage but don't deliver.
But here's what we often forget - humans aren't perfect either. And humans make different kinds of mistakes that we've just gotten used to.
What are the odds you'll catch every missing alt tag across 10,000 images? Pretty slim. How about reviewing the same website six months later and catching new issues that broke accessibility? Even slimmer.
Different accessibility experts test the same site and give different results. Manual audits take forever and cost a fortune. Most websites never get audited at all.
The Game-Changing Difference
Here's where it gets interesting: humans stay roughly the same at accessibility testing. We don't suddenly become 10x better at catching issues or 10x faster at reviewing code.
AI, on the other hand, gets better every single day.
Every time someone corrects an AI's mistake, it learns. Not just for that one website - for every website that uses the same system. When you teach an AI that a particular div pattern should be a button, suddenly thousands of other sites get that fix too.
Human expertise doesn't scale like that. When an accessibility consultant learns something new, only their clients benefit.
What We're Missing While We Wait for Perfect
While we debate whether AI is accurate enough, most of the internet remains completely inaccessible. We're talking about going from maybe 20% of websites being accessible to potentially 80%+ with AI tools that are "only" 95% accurate.
Think about what that means for someone using a screen reader. Right now, they hit broken websites constantly. With widespread AI accessibility tools, even imperfect ones, they'd encounter far fewer barriers online.
The Tesla Moment
People say self-driving cars are too dangerous because they aren't perfect. Meanwhile, human drivers crash constantly due to texting, drinking, falling asleep, or just poor judgment.
Tesla's Autopilot isn't flawless, but it's already much safer than human drivers. And it keeps getting better while human driving skills stay the same.
We're at the same inflection point with web accessibility.
Why Speed Matters More Than Perfection
The accessibility community needs to focus on making AI tools better, faster. Not because they're good enough now, but because they have the potential to solve accessibility at internet scale.
Every month we spend debating AI accuracy is another month that millions of people can't use basic websites. Every correction we make to AI systems today makes them work better for everyone tomorrow.
The question isn't whether we can trust AI to be 100% accurate. It's whether we can afford to wait for perfection while the web remains largely inaccessible.
The Real Opportunity
We're looking at the possibility of the biggest leap in web accessibility in history. Not through better training programs or stricter laws (though those help), but through technology that can actually scale to the size of the internet.
AI accessibility tools that work on 95% of cases across millions of websites will help more people with disabilities than perfect human audits that only cover thousands of websites.
The path forward isn't about trusting AI blindly. It's about improving it rapidly, integrating it into our development workflows, and using it to catch issues that would otherwise never be found.
Bottom Line
AI won't be 100% accurate anytime soon. But it doesn't need to be. It needs to be better than what we have now, and it needs to keep improving fast.
The sooner we embrace imperfect AI tools and start making them better, the sooner we can make the web work for everyone.
If you're working on a tool that uses AI for accessibility audits and want my feedback, I'd be glad to learn about what you're building and provide input.
And if you've already passed the audit phase but are now struggling with accessibility strategy to make your product and webpage accessible by design, contact me at nir@nira11y.com.
Are you working on improving AI accessibility tools? What's the biggest challenge you're facing?