AI Builds Faster. Accessibility Testing Doesn't.

Jamie
ai assisted developmentaccessibility testingwcagautomated testinglanguage access

Jamie · AI Research Engine

Analytical lens: Strategic Alignment

Small business, Title III, retail/hospitality

AI-assisted · Source-linked · Editorially reviewed · Methodology

Trust note

This article was drafted with AI assistance, reviewed against accessibility.chat editorial standards, and should be treated as research and education rather than legal advice. We prioritize primary sources and correct material errors.

Two young adults presenting startup business ideas on a flipchart indoors.
Photo by RDNE Stock project on Pexels

There are no accessibility checkpoints in most AI-assisted pull requests. There are linting rules, formatting checks, security scans, and dependency audits. There is often a detailed CI/CD pipeline with automated tests covering dozens of edge cases. And then there is the question of whether a screen reader user can actually navigate the thing that just shipped—and that question, more often than not, gets asked after the fact, if at all.

This is the gap that Level Access's 2026 State of Digital Accessibility Report (opens in new window) just quantified. The data reveals a structural problem in how teams are using AI, and a clear path forward for teams that want to avoid it.

The Two-Speed Problem

Level Access surveyed more than 2,500 professionals about where AI was accelerating their work across the software development lifecycle. Planning: 58% said AI had sped things up. Design: 61%. Development: 51%. QA and testing? Only 30%.

That asymmetry is the story. AI has compressed the front half of the workflow—ideation, design, code generation—while the validation layer has stayed roughly where it was. The result is a production line where raw material moves three times faster than quality control.

WebAIM's latest automated scan of the top million home pages (opens in new window) makes the downstream consequence visible: automatically detected errors per page rose 10.1% in a single year, from 51 to 56.1. WebAIM named AI-assisted coding as a possible contributing factor. That's not an indictment of AI—it's a description of what happens when you accelerate creation without accelerating verification.

Disabled people encounter those 56 errors. They hit the broken form label, the image with no alt text, the modal that traps keyboard focus. The errors aren't abstract compliance failures—they're barriers to participation. And they're accumulating faster than the industry is catching them.

What Actually Separates the Teams That Are Improving

Here's where the Level Access research gets genuinely interesting, and where cautious optimism is warranted.

The report identified a subset of organizations—the "improvers"—who are getting better across every accessibility metric they track. The researchers then looked for what distinguished them. The answer was not AI adoption. Improvers use AI at roughly the same rate as everyone else.

What separated them was program maturity: training, communication infrastructure, and a culture where accessibility is embedded rather than appended. Critically, professionals in mature programs were nearly six times as likely to say AI helps them prioritize accessibility issues, and nearly three times as likely to say it helps them be proactive in development.

This is the amplifier finding. AI doesn't create accessibility competence—it multiplies whatever competence already exists in the organization. A team with no accessibility context, no testing infrastructure, and no designated ownership will use AI to generate inaccessible code faster. A team with strong foundations will use AI to catch issues earlier, document requirements more thoroughly, and scale their existing judgment.

This connects directly to a pattern documented in accessibility research: automated testing tools—even sophisticated ones—detect at most 37% of real accessibility barriers. AI-assisted code generation doesn't change that ceiling. It just means more code needs to clear it.

Where Human Judgment Has to Live

Andrew Chung, Chief Product Officer at Level Access, frames the solution around three intervention points: finding issues, fixing them, and documenting the evidence. Each one maps to something teams already do—a handoff, a pull request, a review—so the goal is integration, not addition.

The handoff is the earliest leverage point. When a designer passes work to a developer, that's the moment to specify focus order, accessible names, heading structure, and state announcements. Leave those vague, and the AI model fills the gap with its own defaults—which may or may not be accessible. Chung's recommendation: put accessibility specifications where the model actually reads them. Design system components. Rules files. Prompt templates. GitHub's Annotation Toolkit (opens in new window) offers Figma components for marking headings, landmarks, and keyboard behavior directly on the canvas, so intent travels with the file.

The pull request is the second checkpoint. Automated accessibility checks can run in CI/CD pipelines and block or warn on merge. This is where tools like axe-core (opens in new window) and similar automated scanners earn their place—not as the final word on accessibility, but as a first filter that catches the detectable issues before they compound.

The third point is documentation. Human testers need to confirm what automated tools cannot: that focus management works logically, that dynamic content updates are announced correctly, that the experience holds up under real assistive technology. That judgment needs to be recorded, not just performed.

This three-point model responds to the false dichotomy between automated testing and manual audits that has trapped the industry for years. Neither approach alone is sufficient. The question is where each belongs in a workflow that AI has fundamentally restructured.

The Compliance Intersection: Language Access Doesn't Get a Pass Either

One dimension the Level Access report doesn't address—but that compliance teams need to hold alongside it—is language access. Title VI (opens in new window) and the ADA (opens in new window) are enforced by overlapping agencies and affect overlapping populations, but they're almost never audited together.

AI-generated interfaces compound this gap. A codebase built with AI assistance may pass automated WCAG checks and still be entirely inaccessible to a screen reader user who reads Spanish—because ARIA labels, alt text strings, form validation messages, and dynamically loaded content were never translated. Traditional translation services don't touch the accessibility layer. idioma.chat (opens in new window) is a concrete example of infrastructure that does: it translates not just visible text but the full stack of accessibility-relevant strings, including ARIA attributes, modal content, and screen reader announcements. Compliance teams building AI-assisted workflows need to account for both mandates simultaneously, not sequentially.

The Standards Fragmentation Problem Underneath All of This

There's a structural issue that makes the two-speed problem harder to solve than it looks. Organizations aren't navigating one accessibility standard—they're navigating WCAG 2.1, Section 508, EN 301 549, and a growing layer of state regulations, each with different requirements and different enforcement mechanisms. AI tools trained on one standard may generate code that satisfies it while missing obligations under another.

This is the compliance framework paradox applied to AI-assisted development: the faster you build, the more surface area you create for standards fragmentation to produce gaps. Mature programs, as the Level Access data suggests, handle this because they've already built the organizational infrastructure to track multiple requirements. Programs without that infrastructure are building faster toward a more complex compliance problem.

What This Means for Your Workflow Right Now

The Level Access research points toward a practical diagnostic. Take a feature your team shipped recently. Trace it from the first design frame to production. Mark every moment a person made a judgment call about whether it would work for a keyboard user, a screen reader user, or someone browsing at 400% zoom.

If those marks cluster at the end—in QA, in production, in a post-launch audit—you're running the two-speed workflow. You're building fast and validating slow. The fix isn't to slow down the build. It's to move the judgment calls upstream, into the handoff, the pull request, and the design system, where they can travel with the work instead of chasing it.

Intervention PointWhat to CheckPrimary StandardTool/Method
Design handoffFocus order, accessible names, heading structure, state announcementsWCAG 2.4.3 (opens in new window), 1.3.1 (opens in new window)Figma annotation kits, design system specs
Pull request / CIDetectable errors: missing alt text, form labels, color contrastWCAG 1.1.1 (opens in new window), 1.4.3 (opens in new window), 1.3.5 (opens in new window)axe-core, automated pipeline checks
Human reviewFocus management, dynamic content, assistive tech behaviorWCAG 4.1.3 (opens in new window), 2.1.1 (opens in new window)Manual testing with NVDA, JAWS, VoiceOver
Language accessARIA labels, alt text, form validation in all served languagesTitle VI (opens in new window), ADA overlapidioma.chat, full-stack translation infrastructure

The web getting less accessible is not an inevitable consequence of AI-assisted development. It's a consequence of AI-assisted development without the organizational infrastructure to match its pace. The teams proving that out—the improvers in the Level Access data—aren't using AI less. They built the foundation first, and now AI is making their existing judgment faster, not replacing it.

About the Jamie lens

A strategy lens for small business and Title III. Frames findings around cost, sequencing, and what a retail or hospitality operator can realistically act on first.

Jamie is an AI analyst lens, not a human staff member. It helps frame this article through a consistent accessibility perspective.

Specialization: Small business, Title III, retail/hospitality

View all articles using this lens →

Transparency Disclosure

This article was drafted with AI assistance and reviewed against our editorial methodology. We disclose that process so readers can judge the work clearly.