Glossary

A shared vocabulary for the rest of this reference. Terms are grouped into accessibility concepts, AI concepts, the standards and legal landscape, and the handful of ideas that only arise when the two fields meet. Where a term maps to a specific WCAG success criterion, the criterion number is given so you can trace it back to the source.

Accessibility terms

Assistive technology (AT). Software or hardware a person uses to access digital content, such as a screen reader, screen magnifier, switch device, or voice-control system. AI output has to work through assistive technology, not just look right on screen.

Screen reader. Software that converts on-screen content to speech or braille, such as NVDA, JAWS, VoiceOver, and TalkBack. Screen-reader users navigate by structure, moving between headings, landmarks, and lists, so generated content without real semantics is hard to move through.

WCAG (Web Content Accessibility Guidelines). The W3C’s technical standard for accessibility, organized around four principles, which are perceivable, operable, understandable, and robust, often shortened to POUR. The current Recommendation is WCAG 2.2 from October 2023, though much regulation still references 2.1 from June 2018. See WCAG applicability.

Success criterion (SC). A specific, testable requirement in WCAG, identified by a number such as 1.1.1 Non-text Content and a conformance level of A, AA, or AAA. “WCAG 2.1 AA” means meeting all level A and level AA criteria.

Conformance level (A, AA, AAA). WCAG’s three tiers. AA is the near-universal legal and procurement target. AAA is rarely required across a whole product.

ARIA (Accessible Rich Internet Applications). A W3C specification of HTML attributes, including role and the aria-* attributes, that expose the meaning and state of custom UI to assistive technology. Use it to describe semantics that native HTML cannot, and prefer native elements where they exist. The key success criterion is 4.1.2 Name, Role, Value.

Live region. A part of the page, marked with aria-live or with roles like status, alert, and log, whose updates a screen reader announces without the user moving focus there. It is the central mechanism for making streamed and asynchronous AI output perceivable. The related success criterion is 4.1.3 Status Messages, added in WCAG 2.1 AA. See Dynamic and streaming output.

Polite versus assertive live regions. With aria-live="polite" an announcement waits until the screen reader is idle. With aria-live="assertive" it interrupts immediately. Most generated output should be polite. Reserve assertive for genuinely urgent and infrequent messages.

Accessibility tree. The structured representation of a page that the browser exposes to assistive technology, built from the DOM plus ARIA. If information such as a confidence level lives only in CSS or in pixels, it is not in the accessibility tree and assistive technology users cannot perceive it.

Focus management. Controlling which element is active and where keyboard focus moves. When new generated content arrives, focus has to be handled deliberately so the user is neither stranded nor pulled away from what they were reading.

Curb-cut effect. The observation that accessibility features built for disabled people routinely benefit everyone, the way curb cuts help wheelchair users and also people with strollers, luggage, and bikes. It recurs throughout AI accessibility. Structured output, for example, helps screen reader users and also anyone skimming. See Why it matters.

AI terms

Large language model (LLM). A model trained on large amounts of text that generates text by predicting likely continuations. Its output is probabilistic and not authored in advance, which is the root cause of most AI-specific accessibility challenges.

Generative AI. Systems that produce new content such as text, images, audio, and code, rather than only classifying or scoring existing input.

Token. The unit a language model reads and emits, which can be a word, a sub-word, or a fragment of characters. Models generate token by token, which is why output can be streamed as it is produced.

Streaming. Delivering generated output piece by piece as it is produced, rather than all at once. It improves perceived speed for sighted users, but handled naively it floods screen readers with constant announcements. See Dynamic and streaming output.

Inference and runtime. The moment the model actually produces output for a user, as opposed to training time. Because AI content is created at inference, accessibility is a runtime property. See The feedback loop.

Non-determinism. The property that the same input can produce different output across runs. It means you cannot rely on a single fixed DOM that you tested once.

Hallucination. Confidently presented output that is fabricated or incorrect. It is an accessibility concern because the cue that an answer is unreliable has to reach all users, not only those who can see visual hedging. Some researchers argue the term flatters the problem. A widely-cited paper contends that LLM falsehoods are better described as “bullshit,” meaning output produced with indifference to truth, which makes confident wrongness a structural trait rather than an occasional glitch. See Uncertainty and confidence.

Confidence and uncertainty. How sure the model or the system is about an output. Communicating it accessibly, in text and in the accessibility tree rather than only through color or styling, is a distinct design problem.

Prompt and system prompt. The user’s input and the hidden instructions that shape model behavior. A change to either can quietly alter output structure and therefore accessibility.

RAG (retrieval-augmented generation). Adding documents retrieved at query time to a model’s input so it can ground its output in them. Changing the retrieval source can shift tone, structure, and bias without any model change.

Fine-tuning. Training a base model further on additional data to adjust its behavior. Like a model swap or a prompt edit, a fine-tune is an accessibility-relevant change.

Agent and agentic. An AI system that takes actions toward a goal, such as calling tools, navigating, editing, or transacting, sometimes without confirming each step. It raises the question of how users perceive, interrupt, and review actions taken on their behalf. See Agentic actions.

Multimodal. Working across more than one modality, such as text, image, and audio. Multimodal output has to carry equivalent information in an accessible modality, so a generated chart needs a text equivalent.

Standards and legal landscape

EN 301 549. The European harmonized standard for ICT accessibility, with current version V3.2.1 from 2021. It incorporates WCAG 2.1 AA and is the usual reference for demonstrating conformance under EU law.

European Accessibility Act (EAA). Directive (EU) 2019/882. Obligations apply from 28 June 2025 across many consumer products and services. See EU AI Act and accessibility.

EU AI Act. Regulation (EU) 2024/1689, a risk-based regulation of AI systems, in force since 1 August 2024 and phasing in through 2026 and 2027. It is not an accessibility law, but it intersects with accessibility obligations and transparency duties.

ADA (Americans with Disabilities Act). US civil-rights law applied to digital services through case law, and for state and local government through the Department of Justice’s April 2024 Title II rule adopting WCAG 2.1 AA.

Section 508. US law requiring federal ICT to be accessible. Its 2017 refresh references WCAG 2.0 AA.

NIST AI RMF. The US National Institute of Standards and Technology’s AI Risk Management Framework, version 1.0 from January 2023, structured around four functions, which are Govern, Map, Measure, and Manage. It is a voluntary governance scaffold that the evaluation loop maps onto well. See NIST AI RMF.

VPAT (Voluntary Product Accessibility Template). A document reporting how a product conforms to accessibility standards. The completed report is often called an Accessibility Conformance Report, or ACR. It is frequently required in procurement.

Terms specific to AI accessibility

Generated content accessibility. The accessibility of output that no human authored and that is delivered at runtime. It is a property of the generation and rendering system, not of a fixed artifact. See Generated content.

Announcement flooding. The screen-reader failure mode where streamed or rapidly updating output triggers a constant barrage of announcements, making content impossible to follow. You mitigate it by debouncing updates and choosing live-region politeness deliberately.

Provenance. Where an output, or a claim within it, came from. That includes which source, which tool, and whether it is generated or retrieved. It has to be perceivable in the accessibility tree, not implied by visual treatment alone.

Interaction loop and evaluation loop. The two loops that define AI accessibility. The interaction loop is the per-use cycle of perceive, interpret, act, and perceive again. The evaluation loop is the over-time cycle of sample, test, fix, and repeat. See The feedback loop.

Overlay. A third-party script that claims to make a site accessible automatically, often marketed with AI. Accessibility practitioners widely criticize overlays as ineffective and sometimes harmful. It is the recurring cautionary precedent when you evaluate claims that AI can automatically fix accessibility.

AIO (AI optimization). Optimizing a site so AI systems such as chatbots and AI search can read and cite it, an AI-era version of SEO. Proponents argue it overlaps with accessibility because bots and assistive technology hit the same barriers, such as poor structure, heavy JavaScript, and unclear labels, so WCAG conformance helps both. Critics counter that AIO encourages ARIA keyword-stuffing that makes sites less accessible. It is a live debate, not a settled equivalence.

Audience: Engineer

Here is a quick mapping from these terms to the attributes you will actually type. A live region is aria-live="polite" or a role such as status, alert, or log. Name, role, and value come from aria-label or aria-labelledby, the role attribute, and state attributes such as aria-expanded. The accessibility tree is best inspected with your browser’s accessibility inspector or with axe, rather than trusting the rendered pixels. If a concept on this page has no matching representation in the accessibility tree in your implementation, that is the bug.