AI and Accessibility: Useful Accelerator, Not a Compliance Shortcut
Every few weeks another vendor promises that AI will "solve" your accessibility problem. Drop in the tool, flip a switch, and your site is compliant. It's a compelling pitch - and it's wrong in ways that matter legally.
The honest picture is more nuanced. AI is a genuine accelerator for specific, well-defined accessibility tasks. But it does not understand context, cannot judge user intent, and does not bear legal responsibility for the outcome. Under the European Accessibility Act (EAA), your organisation does.
This post maps where AI actually helps in an accessibility programme, where it falls short, and what a defensible human-in-the-loop process looks like in 2026.
Why automation can only ever get you part of the way there
Before evaluating any AI tool, it helps to understand the structural ceiling on automated accessibility testing.
Automated tools - including the most sophisticated AI-assisted scanners - typically catch between 30% and 40% of WCAG violations. That figure comes from multiple independent analyses: TestParty's 2025 review of WCAG testing tools, D2i Technology's 2026 compliance guide, and the practical guidance published by Level Access all converge on the same range. Even axe-core, the open-source engine that powers Google Lighthouse and Microsoft Accessibility Insights, reliably detects around 57% of issues by volume in controlled conditions - and real-world coverage is lower.
What automation misses is not random noise. It is a specific category of failure: contextual and experiential issues. A scanner can flag that an image has an alt attribute. Only a person can judge whether that text actually describes the image meaningfully, conveys its purpose, or should be empty because the image is decorative. The same logic applies to reading order, keyboard flow through complex widgets, screen reader announcements, and cognitive load. As Level Access puts it, automation handles the what while human testers judge the whether it works.
The remaining 43-70% of WCAG issues require manual testing with assistive technologies - keyboard navigation, screen readers such as NVDA and VoiceOver - plus human judgement. A clean automated scan is a starting point, not a finish line.
Where AI genuinely helps
None of this means AI is useless. Used correctly, it compresses real work. Here is where it earns its place:
First-pass alt text at scale. Generating draft image descriptions for a large content library is exactly the kind of repetitive, pattern-based task where AI performs well. Tools can process thousands of images in the time a human team would spend on dozens, giving editors a starting draft rather than a blank field.
Flagging obvious structural issues. AI-assisted scanners reliably catch missing form labels, low colour contrast, absent heading structure, and empty links - the issues that automated rules can express precisely. Running these checks in a CI/CD pipeline means problems are caught before they reach production, not after.
Captions and transcripts for audio/video. AI speech recognition has made captions dramatically cheaper and faster to produce. This matters: video without captions is a WCAG 1.2 failure and an EAA obligation. AI gets you a usable first draft quickly.
Summarising and simplifying content. For cognitive accessibility, AI can help produce plain-language summaries, define jargon, and restructure dense text - supporting users with cognitive or learning disabilities.
Authoring assistance. Accessibility linting integrated into a CMS or code editor catches issues at the point of creation, before content is published. This is prevention rather than remediation.
Improving assistive technology directly. AI is also improving the tools that users with disabilities rely on - better screen reader image descriptions, live captions in video calls, voice control improvements. This is AI helping users, not just developers.
The best use of AI in an accessibility programme is as a first-pass filter and draft generator — not as a decision-maker. It reduces the volume of work that reaches human reviewers; it does not replace them.
Where AI fails - and the risks of over-reliance
The same capabilities that make AI useful also define its limits. These are not edge cases; they are systematic gaps.
AI cannot judge context or purpose. A product image on an e-commerce page, a decorative banner, a chart, and a screenshot of code all require different alt text strategies. AI tools frequently misclassify decorative images as informative (and vice versa), miss text embedded in images, and produce generic descriptions that satisfy the presence check but fail the meaning test. Tools can flag that an image has alt text; only a person can judge if that text actually describes the image well.
The alt text gap is getting worse, not better. A 2026 independent study by webyes retesting 100 EU websites found that missing alt text had risen from 33% to 53% year over year - the most concerning shift in their findings. The study noted that new images are being added to sites without consistent accessibility review. Automation has not closed the gap because new content ships without human oversight. More tools, same problem.
AI-generated captions still need human correction. Real-world studies find platform auto-captions achieve around 60-70% accuracy without human correction - well below the threshold needed for accessibility. AI struggles with accents, rapid speech, overlapping dialogue, technical vocabulary, and tonal cues. WCAG requires captions to accurately convey spoken content; automatic captions are acceptable only when confirmed fully accurate.
AI-generated code can introduce new barriers. The 2026 WebAIM Million report found that increased reliance on AI-assisted coding practices correlates with increased detectable accessibility errors, noting that "vibe coding" trends likely contribute to growing page complexity and ARIA misuse. AI code generators do not reliably produce accessible HTML, ARIA patterns, or keyboard interaction models.
Confident-but-wrong outputs create a false sense of safety. AI tools do not flag their own uncertainty. A generated alt text that is plausible-sounding but contextually wrong will pass an automated check and mislead a human reviewer who is moving quickly. The risk is not that AI produces obviously bad output - it is that it produces subtly wrong output at scale.
"AI accessibility overlays" do not make a site compliant. This deserves its own sentence. Overlay widgets that claim to auto-fix accessibility issues via JavaScript injection use the same underlying detection engines as standalone scanners - achieving 30-40% real-world coverage at best. The FTC explicitly banned accessiBe's "95% compliance" claim in January 2025 as deceptive. For a full analysis of why overlays fall short, see our post Do accessibility overlays make you compliant?
The EAA assigns responsibility to you - not your tools
The EAA (Directive 2019/882) came into force on 28 June 2025; 2026 is the first full year of active supervision across EU member states. Enforcement is no longer theoretical. Audits are being requested, accessibility statements are being scrutinised, and procurement teams are asking vendors for measurable proof of conformance. In June 2026, a French court ordered Carrefour to make its online commerce services fully accessible within six months, with daily fines accruing for each day of delay.
The directive does not care what tools you used. It holds you responsible for the accessible outcome. If your AI-generated alt text is wrong, that is your WCAG 1.1.1 failure. If your AI-captioned video is inaccurate, that is your WCAG 1.2 failure. The tool vendor is not named in the enforcement action.
This is not a reason to avoid AI. It is a reason to build a process where AI output is reviewed by humans before it reaches users - and where that review is documented.
A practical human-in-the-loop workflow
The goal is not to choose between AI and humans. It is to use each where they are strongest.
Integrate axe-core, Lighthouse, or a comparable engine into your build pipeline. Catch missing labels, contrast failures, and structural issues before code merges. This covers the automatable 30–40% consistently and cheaply.
For large image libraries or video backlogs, AI drafts are a genuine time-saver. Treat every output as a draft, not a finished product. Flag images containing text, charts, and infographics for priority human review — these are where AI most frequently fails.
An accessibility specialist reviews AI-generated alt text and captions for context, purpose, and accuracy. This is not a full rewrite of every item — it is a quality gate that catches the contextually wrong outputs that automated checks miss.
Test key user journeys with a keyboard only (no mouse) and with a screen reader — NVDA on Windows, VoiceOver on macOS/iOS. This is the only way to catch reading order issues, keyboard traps, focus management failures, and complex widget behaviour.
Disabled users find issues that neither automated tools nor non-disabled testers identify. This is the highest-signal testing you can do, and it is what regulators and courts increasingly expect to see documented.
Sites change weekly. New content, updated components, and third-party integrations all introduce new barriers. Automated monitoring catches regressions between manual audits. 'We ran an audit last year' is not a strong defence if updates introduced new barriers.
A brief note on the EU AI Act
The EU AI Act (Regulation 2024/1689) becomes fully applicable on 2 August 2026, with transparency obligations under Article 50 taking effect from that date. For teams using AI in accessibility workflows, two things are worth noting.
First, if your AI system interacts directly with users - a chatbot, a live caption system, a personalisation engine - the Act requires clear disclosure that they are interacting with AI, and that disclosure must itself meet accessibility requirements. Second, if AI is embedded in an EAA-covered service (e-commerce, banking, e-books, transport), inaccessible AI UX creates the same legal exposure as any other accessibility barrier. The two regulatory frameworks are converging: both require that digital systems work for everyone, including people with disabilities.
The AI Act does not change your EAA obligations. It adds a parallel layer of transparency and documentation requirements that accessibility teams should be aware of when procuring or deploying AI tools.
The bottom line
AI is a useful part of an accessibility programme. It is not a substitute for one.
The tools that are genuinely worth using - AI-assisted scanners, alt text generators, caption drafters, authoring linters - all share the same characteristic: they reduce the volume of work that reaches human reviewers. They do not eliminate the need for human review.
Under the EAA, you are responsible for the accessible outcome. A human-in-the-loop process - automated scanning, AI-assisted drafting, expert manual audit, assistive-technology testing, and disabled users in your test cohort - is what actually gets you to conformance. That is not a limitation of the technology. It is the nature of accessibility: it is ultimately about whether real people can use your product.
Start with what you can measure. Our free WCAG 2.2 AA checklist covers every Level A and AA criterion in plain English, with pass/fail tests you can run yourself - a practical foundation before you layer in AI tooling or commission a full audit.
Related reading
Accessibility Conformance Reports (ACR/VPAT): The EU Procurement Guide for Vendors and Buyers
What a VPAT is, how it becomes an ACR, which edition to use for EU procurement under EN 301 549, and how to write one that survives scrutiny - plus a buyer-side checklist.
Web Accessibility Directive vs European Accessibility Act: Two EU Laws, One Compliance Programme
The EU has two distinct digital-accessibility regimes: the WAD for public sector bodies and the EAA for private-sector products and services. Many organisations are subject to both. Here's how to tell them apart - and how to satisfy both at once.
Cognitive Accessibility and the EAA: Why WCAG Compliance Isn't Enough
WCAG 2.1 AA is the EAA's legal floor - but it only partially covers people with cognitive and learning disabilities. Here's what teams must do beyond the standard.