NEW: AI Readiness Starts with an Accessible Website
Read the study2026 RESEARCH
The Hidden Risk in AI-Generated Code
AI is helping teams ship code faster than ever, and most trust that code to be accessible. Our survey shows where that trust is misplaced, and what it's costing organizations.
)
On this page
We surveyed 231 web developers, content creators, and business leaders about their use of AI and how accessibility fits into that work.
Seventy-five percent already use AI to write code or create content. Nearly two-thirds say it’s helped them publish more. Most are confident in the accessibility of that output, despite a rise in the number of accessibility issues and complaints.
)
AudioEye ran a companion study to test whether that confidence holds up. We asked five AI tools (OpenAI, Anthropic, Google, xAI, and Lovable) to each build three websites that meet WCAG 2.2 AA requirements, the global standard for accessibility compliance.
All five failed, even though every one of them was told exactly what standard to meet. The average page still had 55 accessibility issues. Ninety-one percent of those issues were medium or high severity, meaning they make tasks harder or impossible for people with disabilities to complete.
That failure is already showing up in legal exposure. Forty-six percent of organizations that use AI to write code or create content have received an accessibility complaint, demand letter, or lawsuit in the past 24 months. More than two-thirds say AI was involved in creating the page in question.
)
Below, we look at three things: how confidence in AI-generated code remains high despite evidence pointing the other way, why general-purpose AI tools can't write accessible code even when prompted, and how the tools organizations use to manage accessibility affect their legal risk.
INSIGHT 1
Confidence in AI is high, even as complaints climb
Eighty-one percent of AI users say they're confident their AI-generated code meets accessibility standards. Yet 73% of that group say accessibility issues or complaints have increased since adopting AI.
INSIGHT 2
Prompting AI for accessibility doesn't produce accessible code
Just 29% of developers say they prompt AI tools to follow accessibility guidelines. Even then, 60% have noticed more accessibility issues since adopting AI. In our own testing, AI tools given that same instruction still averaged 55 issues per page.
INSIGHT 3
Automation alone misses the issues that drive litigation
Seventy-one percent of AI users at organizations that rely on an accessibility widget have received a demand letter or lawsuit in the past 24 months. That's nearly double the rate at organizations that pair automation with human expertise (36%).
Confidence in AI is high, even as complaints climb
Confidence in the accessibility of AI-generated code is high. So is the cost of trusting it without the right expertise in place.
Eighty-one percent of AI users say they're confident the AI-generated code on their site meets accessibility standards. Of those users, nearly three in four say accessibility issues or complaints have increased since adopting AI.
73%
say accessibility issues have increased since adopting AI
That confidence extends to letting AI check its own work. Among AI users, 46% said they would be “very confident” in an LLM’s assessment of their site’s accessibility. Thirty-six percent said they would be “somewhat confident.”
Confidence barely dips across roles, even among those closest to the code.
)
With confidence running that high across every role, there's a higher risk that teams aren't checking their code for accessibility compliance. That gap is already showing up in legal exposure.
Forty-six percent of organizations that use AI to write code or create content have received an accessibility complaint, demand letter, or lawsuit in the past 24 months. More than two-thirds say AI was involved in creating the page in question.
AI helps teams work at scale. But it can't handle accessibility on its own. These models learned to write code from a web that is still mostly inaccessible, so they repeat its mistakes. Without someone who understands accessibility checking that code, confidence runs ahead of reality.
The next section looks at where these tools fall short, and why.
Prompting AI for accessibility doesn't produce accessible code
Just 29% of developers say they prompt AI tools to follow accessibility guidelines. At the same time, 71% of developers say they assume AI-generated code is accessible by default.
)
Even when developers do ask, it barely helps. Sixty percent who specifically prompt AI to follow accessibility guidelines say they’ve seen more issues since adopting AI.
These tools have a hard ceiling, built into how they were trained. General-purpose models learned to write code from the same inaccessible web they're now supposed to fix, not from accessibility experts. An instruction to follow WCAG doesn't undo that training.
We ran our own test to see exactly where that ceiling sits. We asked five leading AI tools to build websites that meet WCAG 2.2 AA, the global accessibility standard. Their pages still averaged 55 accessibility issues each. The average webpage in our 2026 Digital Accessibility Index had 62, so asking for WCAG compliance barely moved the needle. And most of what they missed wasn't minor.
91%
of issues were medium or high severity, meaning they make tasks difficult or impossible to complete and drive the majority of litigation
We aren’t the only ones seeing this. WebAIM's latest report found that the average number of accessibility issues per page(opens in a new tab) went up for the first time in years, and pointed to the rise of AI-generated code as a likely cause.
That same overconfidence carries into how teams review the work. Nearly 40% do no specific accessibility review on AI-generated code at all. One in six don't even know which standard they're supposed to be meeting.
)
Automation alone misses the issues that drive litigation
Every organization that leaves accessibility work to a general-purpose AI tool is exposing itself to risk.
Forty-six percent of AI users have already received an accessibility demand letter or lawsuit. Of those, seven in ten say AI was involved in coding the page in question.
)
That risk lands hardest on organizations that rely only on an accessibility widget. A widget is automation alone. It scans a live site and applies fixes on its own. It never adds a person who understands accessibility to catch the high-risk issues AI-generated code introduces. A more complete approach pairs that same automation with human experts who test with real assistive technology and fix what the automation misses.
Organizations using only a widget are nearly twice as likely to get a demand letter or lawsuit as organizations taking the more complete approach.
71%
of AI users whose organization use only an accessibility widget have received a demand letter or lawsuit
36%
of AI users whose organization uses a more complete approach have received a demand letter or lawsuit
Both groups publish about the same amount of AI-generated work. The real gap is whether anyone is checking it. Widget-only organizations never have a person testing pages with real assistive technology, only the automated tool. Organizations taking the more complete approach back that same automation with real testing from accessibility experts. That combination is likely why they catch more of the issues that drive litigation. It's also likely why they're about half as likely to have received a demand letter or lawsuit.
Closing the gap takes AI built on accessibility expertise
Teams trust AI tools to write accessible code, but the evidence shows they should not. Most AI users are confident their output meets accessibility standards. Most of that same group have seen issues and complaints climb since adopting AI. Almost half of their organizations have already received a demand letter or lawsuit.
Asking a general-purpose AI tool to follow WCAG doesn't get you much further than not asking. These models learned to write code from a mostly inaccessible web, not from accessibility experts. AudioEye's own testing found that AI tools explicitly told to build accessible sites still produced pages nearly as inaccessible as the average website, with 91% of the issues severe enough to make a task difficult or impossible to complete. A model can check code against a rule. It can't test whether someone using a screen reader can actually complete the task. That takes a person trying it.
That gap between writing code and testing it shows up in who gets sued. Organizations relying on automation alone, like a widget, are nearly twice as likely to get a demand letter or lawsuit as those adding accessibility experts. Automation can't catch what it wasn't trained to recognize. People testing with real assistive technology can.
None of this means teams should stop using AI. They just shouldn’t let it run without the right accessibility expertise in place. That means real testing with assistive technology, and using what that testing finds to guide either the AI itself or the people reviewing its output. The teams that build that in now can fix issues on their own schedule. The ones that don’t will keep learning about them in complaints or demand letters.