WCAG applicability to AI

The Web Content Accessibility Guidelines, or WCAG, are the W3C standard for digital accessibility. They are the baseline that almost every accessibility law and contract points to, and they apply to AI products the same way they apply to anything else on the web. An AI feature is still web content. Its buttons, text, images, and status messages all fall under WCAG. Being powered by a model is not an exemption.

Key points

WCAG is the web accessibility standard almost every law points to, and it applies to AI products with no exemption for being AI.
Its four principles are perceivable, operable, understandable, and robust, often shortened to POUR.
WCAG already covers most of an AI product, including text alternatives, real structure, and status messages for streaming.
It does not yet cover uncertainty, streaming volume, agentic actions, or whether generated output is actually any good.
Most law cites WCAG 2.1 AA, while 2.2 is the current Recommendation and a strict superset, so conformance is necessary but not sufficient.

The four principles

WCAG is organized around four principles, often shortened to POUR. Content has to be Perceivable, Operable, Understandable, and Robust. The next four pages take each principle in turn with AI in mind. Here is the short version.

Perceivable. Users can take in the content with whatever sense or assistive technology they rely on. For AI this is mostly about generated images, generated audio and video, and the structure of generated text. See WCAG Perceivable.
Operable. Users can drive the interface with whatever input they have, including a keyboard on its own. For AI this is about stopping generation, moving through long output, and never trapping focus. See WCAG Operable.
Understandable. Users can follow what the content means and can predict how the interface will behave. For AI this is about plain output, predictable behavior despite non-determinism, and signaling when an answer might be wrong. See WCAG Understandable.
Robust. Content keeps working across browsers and assistive technology as they change. For AI this is about valid, semantic markup for generated content and honest use of ARIA. See WCAG Robust.

Which version applies

Three versions of WCAG are in play. WCAG 2.0 came out in 2008, 2.1 in 2018, and 2.2 in 2023. Each one adds success criteria on top of the version before it, so they stack rather than replace.

Most law today references WCAG 2.1 level AA. The European standard EN 301 549, which underpins the European Accessibility Act, incorporates WCAG 2.1 AA. The US Department of Justice adopted WCAG 2.1 AA for state and local government in its 2024 rule under Title II of the ADA. WCAG 2.2 is the current W3C Recommendation and is a strict superset of 2.1, so meeting 2.2 AA means you also meet 2.1 AA. This site holds itself to 2.2 AA for that reason.

A larger rewrite, WCAG 3.0, is in early draft. It would replace the pass or fail A, AA, and AAA model with a score-based outcomes model, where some failures count as critical errors. It is still years from becoming a standard, so plan around WCAG 2.2 today and keep an eye on 3.0.

WCAG has three conformance levels. Level A is the minimum, level AA is the common legal target, and level AAA is rarely required across a whole product. When this site says AA it means all level A and level AA criteria.

What WCAG already covers well for AI

The foundations of this site split an AI product into two surfaces, the interface around the model and the output the model generates. WCAG covers the first surface fully, because it is ordinary web content. It also covers a large part of the second.

A generated image still needs a text alternative under criterion 1.1.1. Generated headings and lists still need real structure under 1.3.1. A response that streams onto the page is a status message under 4.1.3. Much of what makes AI output usable is already required by criteria that exist today. You do not need a special standard to know that a generated chart needs a text equivalent.

Where AI outruns the current criteria

WCAG was written for content that a human author controls. Several parts of AI behavior have no clean success criterion yet.

Uncertainty. No criterion says an interface must tell you when an answer is unreliable. A confidently wrong answer can pass every automated and manual check. See Uncertainty and confidence.
Streaming volume. Live regions are covered, but nothing sets a limit on how often output may update before it floods a screen reader. See Dynamic and streaming output.
Agentic actions. WCAG assumes the user acts and the page reacts. It says little about a system that acts on its own, and about how a user reviews or undoes those actions. See Agentic actions.
Output quality. WCAG can check that a text alternative exists and is exposed to assistive technology. It cannot check that an AI-written description is accurate or right for the context. See Generated content.

These gaps are the reason this site exists. WCAG conformance is necessary. On its own it is not enough to prove that an AI product is usable.

The standards being built for the gap

Work is underway to cover the parts WCAG does not reach. The W3C has an AI Accessibility draft that maps AI across authoring tools, user interfaces, evaluation, and assistive technology. Canada has published CAN-ASC-6.2:2025, a standard for accessible and equitable AI systems framed around outcomes rather than fixed technical rules. For the regulatory and governance side, see EU AI Act and NIST AI RMF.

Audience: Accessibility Specialist

Be precise about which WCAG version the obligation actually cites. The EAA and most national law point at 2.1 AA through EN 301 549, while the current Recommendation is 2.2. Testing against 2.2 is good practice, but the legal minimum in the EU is still 2.1 AA until EN 301 549 is updated. The deeper issue is scope. For a generative feature you are not auditing one page, you are sampling a population of possible outputs. Say how you sampled, test the rendering pipeline rather than one frozen transcript, and remember that 1.1.1 checks whether a text alternative exists, not whether it is any good.

Audience: Product Manager

The line to carry into planning is that WCAG AA is the floor, not the ceiling, for AI. A feature can pass every WCAG check and still fail a blind user, because the model was confidently wrong or because the answer flooded their screen reader. Treat WCAG conformance as the entry requirement and budget separately for the AI-specific gaps around uncertainty, streaming, and agent actions.

Audience: Engineer

The four gaps map to concrete work. Uncertainty means exposing confidence in the accessibility tree, not only in styling. Streaming means a debounced live region strategy. Agentic actions mean keyboard-operable review and undo. Output quality means guiding and checking generation rather than trusting it. The WCAG criteria you already know, such as 1.1.1, 1.3.1, 4.1.2, and 4.1.3, still carry most of the load.