Every few months someone asks us, half-joking, whether AI is about to put accessibility auditors out of a job. Usually it's a founder who just watched a demo where a chatbot scanned a homepage in eleven seconds and spat out a tidy list of contrast failures. It looks like magic. It looks like the whole job just got automated.
Here's the honest answer, and it isn't a hedge. It's a real, testable line: AI is already excellent at the half of this job that is pattern-matching against a rulebook. It is still unreliable at the half that requires actually being the user - sitting inside a screen reader, moving through a page in the order a real person experiences it, and knowing whether what you just heard actually made sense.
Both halves are real work. Only one of them is close to solved.
The Stat: Deque's own published research puts automated accessibility scanning at catching roughly 57.38% of WCAG issues on its own - meaning a substantial share of real issues require a human tester by design, not by current tooling limitation alone. (Source: Deque)
The job splits into two halves, not one
The mistake most people make is treating "accessibility audit" as a single task. It isn't. It's two very different jobs wearing one job title.
The first half is rule-checking. Does this image have an alt attribute? Does this text meet a 4.5:1 contrast ratio against its background? Does this form input have an associated label in the DOM? These are yes/no questions with objectively correct answers, and a machine can answer millions of them per second. This is exactly the kind of work automated scanners and AI models were built for, and they've gotten genuinely good at it.
The second half is judgment. Does this alt text actually describe what a sighted user would understand from the image, or does it just technically exist? Does the page make sense when read in the order a screen reader encounters it, rather than the order it looks correct visually? Can a real screen reader user, using a real keyboard, actually complete checkout without getting stuck in a modal that traps focus? These aren't yes/no questions. They require understanding intent, context, and consequence - three things current AI models still struggle to reliably supply.
Where AI is already doing real work
This isn't a "AI is useless" article. That would be dishonest in the other direction, and it isn't what the data shows.
Research covered in Deque's published research on automated accessibility testing coverage confirms that automated tooling has a real, growing footprint in catching legitimate WCAG issues at scale, fast enough to run on every deploy. And a recent look at LLM-generated code, the CHI 2026 study referenced above, found that models perform well specifically on simple, well-defined checks like contrast and alt-text presence. That's not a small thing. It means teams can catch a large volume of low-hanging, high-frequency issues before a human auditor ever opens the page, which is exactly how a modern audit workflow should work: automation for coverage and speed, humans for the parts that need a brain in the loop.
The same WebAIM Million report has documented, year over year, that the most common failures on the web are the boring, structural ones: missing alt text, low contrast, unlabeled form fields. Those are precisely the failures AI is best at catching. If that were the whole story, the "AI replaces auditors" narrative would be correct.
Where it still falls short - and why that matters more than it sounds
It isn't the whole story, because the same research that shows AI succeeding at simple checks shows it degrading sharply on judgment-heavy patterns, like complex ARIA implementations where the "correct" answer depends on what the component is actually supposed to do, not just what markup is present.
The clearest evidence of this gap isn't theoretical. A CHI 2026 paper testing AI-generated image descriptions found that they actually reduced real-world screen reader task completion rates compared to human-written descriptions. Not because the AI descriptions were wrong in a technical sense - they existed, they were grammatically fine, they passed the "is there an alt attribute" check. They failed because they described what was in the image rather than what mattered in the image for the task the user was actually trying to complete. A human writing alt text for a checkout button knows to say "Complete purchase - $84.20" instead of "a gray rectangular button." An AI model, absent that context, often doesn't.
We saw this same pattern up close in our own composite test of 5 AI overlay tools: every tool we tested could point at surface-level markup issues, and every single one failed at the parts of the job that required understanding what a real assistive-technology user was trying to accomplish on the page.
A practical way to think about the split
| Task | Reliable today with AI/automation | Still needs a human tester |
|---|---|---|
| Detecting missing alt attributes | Yes | - |
| Measuring contrast ratios | Yes | - |
| Detecting unlabeled form fields in markup | Yes | - |
| Judging if alt text is actually meaningful | Partial, unreliable | Yes |
| Verifying logical reading order | No | Yes |
| Confirming a screen reader user can complete a real task | No | Yes |
| Reviewing complex ARIA widget behavior | Partial, unreliable | Yes |
Use this as a working checklist when you're deciding what to trust an automated tool to catch on its own, and what still needs a person behind the keyboard:
- Run automated scans on every build for contrast, alt-attribute presence, and form labeling
- Treat automated "pass" results as a floor, not a certificate of accessibility
- Have a human test actual task completion with a screen reader on critical flows (checkout, signup, search)
- Have a human review whether alt text and ARIA labels convey the right meaning, not just a meaning
- Re-test reading order and focus order by hand after any layout or component change
- Never present an automated-only scan as a compliance audit
The honest bottom line
AI hasn't replaced accessibility auditors, and the data so far suggests it isn't about to. What it has done is take over the repetitive, rule-based half of the job, which is genuinely useful, because it frees human testers to spend their time on the half that actually requires being a person: judgment, context, and real task completion. The teams getting this right aren't choosing AI or humans. They're using automation for coverage and speed, and humans for everything automation still can't see.
If you want to know exactly where that line falls on your own site, that's the easiest way to find out. Get the combined automated-plus-human audit and see, in plain language, what a scanner catches on its own and what only a real tester would have found.