Can AI fix web accessibility, or is it making the problem worse?
People complained. Regulators stepped in. Deadlines got extended. The AI-will-fix-everything narrative around digital accessibility compliance has a serious credibility problem, and it earned every bit of it. But buried inside that mess, some things genuinely are working. Here’s the honest picture for 2026.
AI cannot auto-fix WCAG 2.1 AA compliance at scale, not reliably, not yet. But it is genuinely solving a narrow set of specific accessibility problems. The industry has been blurring that line. This blog separates them.
The web has had an accessible digital content problem for decades. Roughly 1 in 6 people globally lives with some form of disability (WHO, 2023), yet most websites, apps, and digital tools were built as if they didn’t exist. Alt text was missing. Videos had no captions. Forms broke screen reader compatibility. Contrast ratios were afterthoughts.
When AI entered this space, the pitch was seductive: drop a script onto your site, let the algorithm handle WCAG compliance automation, done. Organisations bought it. The algorithm, as it turned out, was not done. Neither was the DOJ.
Why the DOJ Extended the ADA Title II Deadline
In 2024, the U.S. Department of Justice finalised new ADA Title II accessibility rules requiring state and local government websites and apps to meet WCAG 2.1 AA standards. The compliance timeline, however, was extended, and the DOJ’s own Interim Final Rule was explicit about why.
The DOJ cited four interconnected reasons in its ADA Title II deadline extension. The 2026 Interim Final Rule was explicit: no single cause, but four compounding ones.
The DOJ acknowledged it had overestimated the readiness of AI-driven remediation tools and web accessibility automation broadly, a sector-wide gap between what vendors claimed and what tools actually delivered.
Genuine WCAG 2.1 AA compliance, when done properly, was more expensive than originally projected, particularly for smaller government entities with limited digital budgets.
Many public sector organisations lacked the internal expertise to implement, verify, or maintain accessible digital content without significant outside support.
With compliance standards still being interpreted by courts, the litigation landscape created real uncertainty that complicated rollout planning and timelines.
The technology concern is most relevant to this blog's thesis, but the DOJ did not single out any one product category as the sole cause. The accessibility overlay market became the most visible example of a broader pattern of over-promise.
Accessibility overlays, JavaScript widgets dropped onto websites claiming to auto-fix WCAG violations, became the most visible symbol of AI over-promise in this space. More than 800 accessibility professionals signed an open statement against them. The National Federation of the Blind filed formal complaints. Blind users published detailed accounts of overlay tools actively conflicting with their own assistive technology. The DOJ’s broad technology concern and the overlay controversy are related but distinct. The overlays are the sharpest, most documented example of a broader pattern of WCAG compliance automation being sold as a complete solution when it was not.
What Users Were Actually Experiencing
Behind the regulatory language were real people hitting real walls. Here is what the complaints documented across court filings, advocacy statements, and published user research:
AI overlays broke existing screen reader compatibility. Users relying on JAWS or NVDA found that overlay scripts intercepted keyboard commands, re-read page elements out of order, or hid content that the user's own assistive technology had already handled, creating a worse experience than no overlay at all.
Source: AccessiBe lawsuit filings; National Federation of the Blind statements, 2021–2024AI-generated alt text accuracy was unreliable at scale. Auto-generated descriptions labelled photos of wheelchair ramps as "people on stairs," and described data charts as generic "image of graph." Confidently wrong descriptions are worse than missing ones for a screen reader user navigating a form or shopping page.
Source: WebAIM accessibility survey data; documented user reportsAI sign language avatars misread facial grammar. In sign language, facial expressions carry syntactic, not just emotional, meaning. An eyebrow position changes a statement into a yes/no question. AI avatars generating sign language were regularly producing grammatically broken output, a fact documented by Deaf community researchers and linguists.
Source: Deaf community accessibility research; British Sign Language studiesCloud AI tools raised data privacy concerns. Screen readers and accessibility tools relying on cloud AI to process page content were transmitting sensitive user interactions, including medical form entries, government document content, to third-party servers, often without adequate disclosure.
Source: EPIC privacy advocacy filings; accessibility community documentationDisability was systematically underrepresented in AI training data. Research has documented that disability-related objects and scenarios, including braille devices, mobility aids, atypical physical environments, appear far less frequently and are recognised less accurately in major image datasets. Models trained on biased data produce biased outputs, and disabled users absorbed that cost directly.
Source: Microsoft Research on dataset representation gaps; AI fairness literature"AI is not fixing accessibility" and "AI is fixing specific accessibility problems" are both true at the same time. Pretending otherwise serves no one, least of all the 1.3 billion people this is actually about.
AI Accessibility Tools That Are Actually Working
Once you separate the over-promised web accessibility automation market from the rest, a smaller, more credible set of real-world wins emerges. These are not cures. They are specific problems being addressed by specific AI accessibility tools, with measurable outcomes and honest caveats.
Be My Eyes + Microsoft: 90%+ Call Resolution
Blind Microsoft customers calling the Disability Answer Desk get GPT-4 vision support via Be My AI, reading error messages, describing interface states, walking through software steps. Over 90% of calls are resolved without a human agent. This works because it is a contained, high-context use case with a defined success metric.
Axe DevTools AI: Violations Caught at Code-Time
Deque's AI toolkit flags WCAG 2.1 AA violations in real time as developers write code, including contrast failures, missing ARIA labels, keyboard traps, before they ship. The key distinction from overlays: it fixes problems at the source, not patches them post-launch. AI surfaces the issue; the accessibility specialist makes the call. That division of labour is what makes it work.
Jamworks: AI Captions Accuracy in Live Lectures
Combining speech recognition with generative AI to transcribe and organise lectures in real time, Jamworks was piloted across UK further education colleges. Hearing-impaired students described it as a genuine accommodation, not a workaround. This is one of the cleaner examples of AI captions accuracy applied to a bounded, real-world problem.
Microsoft + Be My Eyes: Fixing Biased Training Data
Rather than patching apps, this partnership fixes AI at the root by feeding real-world video from blind users into Microsoft's model training pipeline so AI learns disability-context scenarios it previously missed. Fixing dataset representation is slow, foundational work. It is also the most durable kind of fix.
Speech Accessibility Project: Training on Atypical Voices
The University of Illinois initiative gathers voice data from people with ALS, cerebral palsy, stuttering, and other conditions to train voice recognition that actually works for them. Most voice AI was built on "standard" speech, locking out the people who most need hands-free technology.
AI Wearables: Real-Time Environmental Description
AI-embedded glasses now provide real-time audio descriptions of physical environments, reading store labels, narrating museum exhibits, flagging obstacles on streets. These work because they are scoped to a specific task: describe what the camera sees, accurately, out loud. The use case is defined; the feedback loop is direct.
"AI is not solving accessibility. It is making progress on specific, high-leverage problems for the first time at scale, and those are not the same thing."
— The distinction the industry keeps blurring, and regulators are now enforcingWhy the AI vs Human Accessibility Testing Debate Matters for Compliance
When vendors blur the line between “AI helps with some accessibility tasks” and “AI solves accessibility,” real harm follows. Organisations stop investing in actual accessibility remediation expertise. Developers stop learning WCAG. Disabled users get handed a broken overlay and told the problem is solved, until the lawsuit arrives.
The DOJ’s extended timeline was not leniency. It reflected a finding, shared by the advocacy community, the courts, and disabled users themselves, that compliance built on tools that don’t reliably work is not compliance. It is liability deferred.
The European Accessibility Act 2025 is now in enforcement. For companies operating in or selling into EU markets, the stakes are concrete: the global spending power of people with disabilities is estimated at $8 trillion annually (Return on Disability, 2023). That is the market being addressed, or excluded, by decisions about digital accessibility today. “We deployed an AI overlay” is not going to satisfy a European regulator any more than it satisfied the DOJ.
Here is where AI genuinely stands, not where vendors say it stands. This is the distinction between what AI accessibility tools are actually doing versus what is still a marketing claim.
AI Is NOT Reliably Doing
- Auto-fixing entire websites to WCAG 2.1 AA compliance
- Generating accurate alt text at scale without human review
- Producing grammatically correct sign language
- Replacing human accessibility audits or VPAT documentation
- Working reliably with atypical speech or edge-case visual scenarios
- Guaranteeing screen reader compatibility across assistive technology
AI IS Making Real Progress On
- Catching WCAG violations at code-time before they ship
- Visual assistance in bounded, high-context customer service scenarios
- Real-time AI captions accuracy in structured educational environments
- Describing physical environments via wearables within defined scope
- Fixing biased training data representation at the model level
- Expanding voice recognition to atypical and diverse speech patterns
The Path Forward: Multiplier, Not Replacement
The credible trajectory for AI in accessibility is as a tool that accelerates human expertise, not one that automates it away. The implementations that work pair AI with specialists: AI flags likely violations at scale, human experts confirm and prioritise. AI generates draft alt text, content teams review. AI transcribes live, humans correct for high-stakes contexts. That framing is less exciting than “AI fixes everything.” It is also what the evidence supports.
The DOJ’s involvement, the advocacy community’s documented pushback, and the formal complaints are not obstacles to AI-driven digital accessibility compliance 2026. They are the feedback loop that may finally force the industry to build something reliable rather than something marketable. Given that the problem affects over a billion people globally, that correction is overdue.
More on ADA, Section 504, and Digital Accessibility Compliance
This blog is part of our ongoing series on digital accessibility law, AI limitations, and what WCAG 2.1 AA compliance actually requires in 2026. Share a thought or explore the related reads.
