For six consecutive years, web accessibility improved. The gains were modest, but they moved in the right direction every year from 2019 through 2025.

In 2026 the trend reversed.

The WebAIM Million found detectable WCAG 2 failures on 95.9 percent of the top one million home pages, up from 94.8 percent the year before. Average detected errors per page rose 10.1 percent to 56.1. The same three problems have topped the list for seven straight years: low contrast text on 83.9 percent of pages, missing alternative text on 53.1 percent, and missing form labels on 51 percent.

WebAIM's own explanation for the reversal is one the industry will not enjoy. They point to "increased reliance on 3rd party frameworks and libraries and automated or AI-assisted coding practices," alongside page elements growing 22.5 percent in a single year and ARIA attributes rising 27 percent to more than 133 per page.

More markup, more complexity, and more generated code produced worse results. The first year that AI-assisted production went properly mainstream is the year the line bent the wrong way.

The design system decides which way this goes

DORA's 2025 research describes the mechanism without naming accessibility. Teams with quality infrastructure find that AI acts as a powerful collaborator. Teams with "fragmented tooling, siloed data, or fragile infrastructure" find that AI "will simply help them generate technical debt faster."

For design organizations, that infrastructure is the design system.

A team with a mature system has accessible components, defined states, documented interaction behavior, and clear review ownership. Generation flows through that structure and comes out largely conformant, because the constraints are built into what gets produced.

A team without one has a component library that drifted from production two years ago, undocumented usage rules, and accessibility handled by whoever happened to remember. Generation flows through nothing, and the output arrives fast, plausible, visually competent, structurally wrong, and in far greater quantity than anyone can review.

The size of that difference has been measured. Guriță and Vatavu, presenting at the 2025 Web for All Conference, compared UI code generated without any accessibility direction against the same tasks with accessibility explicitly specified. Expert evaluation found a 58 percent violation rate in the unspecified condition against 19 percent when accessibility was named. Under agnostic prompting, alternative text and information structure failed 100 percent of the time, and keyboard navigation failed 80 percent of the time. Naming the requirement took keyboard accessibility to zero violations.

Read that as an organizational finding rather than a prompting tip. It means the accessibility of everything your organization ships now rests on whether the person who opened the tool that morning thought to name the requirement. A design system is how you stop depending on that, because it moves the constraint out of individual discipline and into infrastructure.

The gap between those two teams predates AI. What AI changed is throughput, and it multiplied whichever situation an organization was already in.

Production capacity used to do your quality control for free

Traditional design workflows had a natural governor. A designer produced a handful of concepts, the team reviewed them, and a smaller number survived. Scarcity did quality control at no cost.

That governor is gone. A product manager can generate five interface variations to explain a requirement. A developer can prototype to test feasibility. A marketer can produce a landing page variant. None of them are doing anything wrong, and most of that work is genuinely useful.

The output does have a consistent signature, though. Nielsen Norman Group's evaluation of AI design tools against real project scenarios found flat and interchangeable visual style, generic aesthetics with no brand differentiation, related elements separated by excessive spacing, absent visual hierarchy, and patterns imported from entirely the wrong context. In one case, the tools applied social-media profile conventions to a learning dashboard and promoted secondary information over primary.

Multiply that signature across everyone in the organization who can now produce an interface, and consistency stops being a craft concern and becomes an operational one.

What a design system needs that most do not have

Most design systems are visual libraries with implementation notes. They show what a component looks like and roughly how to build it. That was sufficient when the only consumers were designers and developers who could ask a colleague.

Generation tools cannot ask a colleague. To use a component correctly rather than merely reproduce its appearance, the system has to encode what an experienced person would have known implicitly:

• What customer need the component serves, and when it is the wrong choice
• Required content, character limits, and what happens when content exceeds them
• Every state, including loading, empty, error, partial, and permission-restricted
• Keyboard interaction and screen reader behavior described specifically, rather than promised in general
• The business rules that constrain it
• Related components, and how to choose between them
• The implementation mistakes that keep recurring

This information helps people at least as much as it helps tools. Most design systems have never written it down, because the people who knew it were reachable on Slack. Writing it down is the work, and it is the highest-leverage design system investment available in 2026.

The compliance floor moved while you were adopting tools

European Accessibility Act enforcement began June 28, 2025. The harmonized standard, EN 301 549 v3.2.1, incorporates WCAG 2.1 Level AA in full, and a version covering WCAG 2.2 is expected this year. Coverage extends to e-commerce, banking, electronic communications, and audiovisual media, and it applies to any organization selling to EU consumers regardless of where that organization is headquartered. National penalties run to roughly €900,000 in Sweden and €600,000 in Spain and reach up to 5 percent of annual turnover for serious breaches in France and Italy.

That leaves most large organizations in an awkward position: production volume rising sharply, measured accessibility declining for the first time in six years, and enforcement now live.

A design system with accessibility encoded into its components is the only control that scales alongside the volume. Manual review does not scale, and automated checking has a documented ceiling, catching 57.38 percent of accessibility issues across only 16 of the 50 WCAG success criteria, according to Deque's analysis of more than 13,000 pages. No amount of inspection covers the gap those two limits leave, which is what makes constraint the only workable control.

AI helps maintain the system too

The relationship runs both directions, and this second direction is the one design system teams under-exploit.

Design system maintenance is largely comparison work at a scale humans find tedious: finding duplicate components, identifying where production has drifted from the standard, spotting outdated guidance, drafting documentation, organizing contribution requests, and preparing release notes. That is precisely the shape of work AI does well, and design system teams are chronically under-resourced for it.

Deciding what becomes a standard stays with the humans who own the system. The evidence-gathering that informs the decision does not have to.

Measuring a design system when volume is the variable

Component count was never a good metric, and it is actively misleading now. What matters is whether the system is governing production:

• Percentage of shipped work using approved components
• Percentage of AI-generated concepts that pass review without structural rework
• Accessibility issues found in production versus caught before release
• Design and development rework attributable to inconsistency
• Adoption outside the design team, across product, engineering, and marketing

That last measure is the leading indicator. The design system's job has changed from serving designers to constraining everyone who can now generate an interface. If product managers and developers are not using the system, it is not doing the job the organization currently needs done.

What this year's data says

The organizations that will get the most from AI-assisted design are the ones that did the unglamorous infrastructure work first: documented components, encoded accessibility, clear ownership, and a system that matches what is running in production.

That work has no launch moment and no demo. On this year's evidence, it is also the difference between AI improving your customer experience and AI degrading it at speed.

The 2026 numbers are the first real evidence of which of those outcomes is more common, and the answer does not flatter the industry.

Schedule an AI-Enabled Experience Design Working Session Learn more

Sources

• Alexandra-Elena Guriță and Radu-Daniel Vatavu, "When LLM-Generated Code Perpetuates User Interface Accessibility Barriers, How Can We Break the Cycle?" Proceedings of the 22nd International Web for All Conference (W4A '25), April 2025. https://dl.acm.org/doi/10.1145/3744257.3744266
• WebAIM, "The WebAIM Million," 2026 report. https://webaim.org/projects/million/
• DORA, "Balancing AI tensions: Moving from AI adoption to effective SDLC use," 2025. https://dora.dev/insights/balancing-ai-tensions/
• Deque, "The Automated Accessibility Coverage Report." https://www.deque.com/automated-accessibility-coverage-report/
• Level Access, European Accessibility Act compliance overview. https://www.levelaccess.com/compliance-overview/european-accessibility-act-eaa/
• Huei-Hsin Wang and Megan Brown, "Good from Afar, But Far from Good: AI Prototyping in Real Design Contexts," Nielsen Norman Group, October 24, 2025. https://www.nngroup.com/articles/ai-prototyping/