# Why it matters Roughly 1.3 billion people, about 16 percent of the world's population, live with a significant disability, according to the [World Health Organization](https://www.who.int/news-room/fact-sheets/detail/disability-and-health). AI is moving from novelty to infrastructure. It is embedded in search, email, customer support, hiring, healthcare, and government services. As that happens, the question of whether disabled people can use these systems stops being a niche concern and becomes a question about access to ordinary life. There are three overlapping reasons the answer has to be yes. It is the right thing to do, the law increasingly requires it, and building for it produces better products for everyone. ## Key points - A large share of the world's population lives with a significant disability, and AI is becoming infrastructure they need to use. - There are three reasons to act. It is the right thing to do, the law increasingly requires it, and it makes better products for everyone. - Accessibility law has moved from encouraged to required, and AI products are in scope. The European Accessibility Act applies from June 2025. - Generative AI can produce biased or demeaning output, partly because varied disability data is often excluded from training as outlier data. - Accessible AI features help everyone, but access is worth building even when it only helps disabled users. ## The human case Accessibility is fundamentally about not excluding people from things they have a right to take part in. When an AI system becomes the front door to a service, inaccessibility is not a degraded experience, it is a locked door. That system might be the only way to reach support, the tool that screens job applications, or the assistant that summarizes your medical results. Anna E. Cook frames accessibility as [digital infrastructure](https://annaecook.com/writing/2026/2/2/accessible-design-is-digital-infrastructure) rather than a finishing touch. Infrastructure is invisible until it breaks, and accessibility is often where the break shows first. It also erodes quietly when maintenance stops being a priority. When a platform cuts the people who maintained its accessibility, the features disabled users relied on degrade with no announcement, which is one reason this is continuous work and not a one-time build. See [the feedback loop](https://artificia11y.ds.house/foundations/feedback-loop/). The population is broader than many teams assume. It includes permanent disabilities such as blindness, deafness, and motor and cognitive disabilities. It includes temporary ones such as a broken arm or an eye infection. It includes situational ones such as bright sunlight, a noisy room, or holding a child. It also includes the large and growing share of older adults who acquire impairments with age. Design that assumes a single typical user, one who sees the screen, reads quickly, uses a mouse, and processes generated text without difficulty, excludes a substantial minority outright and serves the rest worse than it could. Generative AI raises the stakes in a specific way, because it produces language and decisions, not just interfaces. A system that generates patronizing, biased, or inaccessible output does a particular kind of harm to the people it misrepresents or talks down to. The bias is measurable, not anecdotal. Penn State researchers found that NLP sentiment and toxicity models score statements about disability as [significantly more negative and toxic](https://www.psu.edu/news/information-sciences-and-technology/story/trained-ai-models-exhibit-learned-disability-bias-ist) than equivalent statements, with the word "blind" flagged as negative regardless of meaning. The root cause, as researchers at Harvard [have argued](https://news.harvard.edu/gazette/story/2024/04/why-ai-fairness-conversations-must-include-disabled-people/), is that disability data is so varied that it gets treated as outlier data and excluded from training, so the people most affected are designed out from the start. See [Bias and representation](https://artificia11y.ds.house/inclusive-development/bias-and-representation/). ## The legal case The regulatory environment has shifted decisively from encouraged to required, and AI products are squarely in scope. - **European Accessibility Act (EAA).** This is [Directive (EU) 2019/882](https://eur-lex.europa.eu/eli/dir/2019/882/oj). Member-state obligations apply from 28 June 2025, covering a broad range of consumer-facing products and services such as e-commerce, banking, e-books, and transport ticketing. Many AI-powered services fall within its scope, and conformance is generally demonstrated against [EN 301 549](https://www.etsi.org/deliver/etsi_en/301500_301599/301549/03.02.01_60/en_301549v030201p.pdf), which incorporates WCAG 2.1 level AA. See [EU AI Act and accessibility](https://artificia11y.ds.house/guidelines/eu-ai-act/). - **EU AI Act.** This is [Regulation (EU) 2024/1689](https://eur-lex.europa.eu/eli/reg/2024/1689/oj), in force since 1 August 2024 and phasing in through 2026 and 2027. It is a risk-based product-safety regulation rather than an accessibility law, but it intersects with accessibility through transparency duties and through high-risk classifications for uses like employment and access to essential services. It explicitly references accessibility requirements aligned with the EAA. - **United States.** The [Americans with Disabilities Act](https://www.ada.gov/) has been applied to digital services through litigation for years. The Department of Justice's [April 2024 Title II rule](https://www.ada.gov/resources/2024-03-08-web-rule/) adopts WCAG 2.1 AA as the technical standard for state and local government web content and mobile apps. [Section 508](https://www.access-board.gov/ict/) requires federal ICT to meet WCAG 2.0 AA through its 2017 refresh. Neither carves out AI. - **Elsewhere.** Many jurisdictions reference WCAG by adoption, including [Canada](https://laws-lois.justice.gc.ca/eng/acts/A-0.6/), [Australia](https://www.legislation.gov.au/C2004A04426/latest/text), and the UK through the [Equality Act](https://www.legislation.gov.uk/ukpga/2010/15/contents) and [PSBAR](https://www.legislation.gov.uk/uksi/2018/952/contents/made). The common thread is that WCAG conformance is the de facto legal baseline almost everywhere, and being AI is not a recognized exemption. AI-specific accessibility standards and oversight are now arriving on top of this baseline. Canada published [CAN-ASC-6.2:2025](https://accessible.canada.ca/creating-accessibility-standards/asc-62-accessible-equitable-artificial-intelligence-systems/6-introduction), a national standard for accessible and equitable AI systems framed around outcomes, which are equitable benefit, the avoidance of inequitable harm, and preserved user agency. The European Disability Forum maintains a [guide to monitoring the AI Act](https://www.edf-feph.org/publications/a-disability-inclusive-artificial-intelligence-act-updated-guide-to-monitor-implementation-in-your-country/) from a disability-rights perspective, and the US Access Board has [held hearings](https://www.access-board.gov/news/2024/07/09/u-s-access-board-holds-hearings-on-artificial-intelligence-ai-for-disability-community-and-ai-practitioners/) on AI and disability. > [!ACCESSIBILITY-SPECIALIST] > > One nuance is worth getting right in conformance work. The EAA and most national laws point at [WCAG 2.1](https://www.w3.org/TR/WCAG21/) AA through [EN 301 549](https://www.etsi.org/deliver/etsi_en/301500_301599/301549/03.02.01_60/en_301549v030201p.pdf), whose current version is V3.2.1 from 2021, whereas the latest WCAG Recommendation is [2.2](https://www.w3.org/TR/WCAG22/) from October 2023. This site tests itself against 2.2 AA, which is a strict superset, but the legal minimum in the EU is still 2.1 AA until EN 301 549 is updated to reference 2.2. When you scope an audit, be explicit about which version the obligation actually cites, because the answer changes which success criteria are mandatory rather than aspirational. The harder gap is that none of these standards yet has criteria purpose-built for generated, probabilistic output, so conformance to WCAG is necessary but not sufficient evidence that an AI product is usable. > [!PRODUCT-MANAGER] > > Three risks are worth raising in a roadmap conversation. > > - **Legal exposure.** For EU-facing consumer services the EAA obligations are live as of June 2025, and remediation under deadline costs far more than building it in. > - **Procurement.** Public-sector and large-enterprise buyers increasingly require a VPAT or accessibility conformance report, and not having one removes you from deals before the conversation starts. > - **Reputation.** An AI assistant caught producing inaccessible or demeaning output is a recognizable and screenshot-able failure. > > The defensible position is to treat accessibility as a launch gate with the same status as security or privacy, not a fast-follow. ## The product case The [curb-cut effect](https://artificia11y.ds.house/foundations/glossary/#accessibility-terms) is real, and AI amplifies it. Features built to make generated content accessible tend to improve the product for everyone. - **Structured output**, meaning real headings, lists, and semantic markup, is navigable by a screen reader and is also easier for anyone to skim and for downstream systems to parse. - **Clear state communication**, such as generating, done, and stopped, is something a blind user relies on, and it also reassures a sighted user that the system has not frozen. - **Honest uncertainty signaling** that a screen-reader user needs to hear also protects every user from over-trusting a confidently wrong answer, which is a safety property and not only an accessibility one. - **Interruptible, reviewable agent actions** designed so a disabled user stays in control are exactly the guardrails that prevent costly mistakes for all users. Inaccessible AI, by contrast, fails quietly and broadly. The same wall of unstructured streamed text that breaks a screen reader also exhausts a user with a cognitive disability and annoys everyone else. Accessibility problems in AI are rarely only accessibility problems. They are usually general quality problems that disabled users hit first and hardest. There is a newer product reason worth naming. The same structured markup that makes a page accessible is also what makes it legible to machines, including the AI systems teams are now racing to build on. Anna E. Cook [makes this case directly](https://annaecook.com/writing/2026/ai-doesnt-fix-accessible-systems-it-depends-on-them), that accessible systems are parseable systems, so the semantic integrity a screen reader relies on is the same integrity a model relies on to interpret a page. Skipping that work does not just exclude disabled users. It removes the foundation the AI itself needs. And the cost is now measurable. The 2026 [WebAIM Million](https://webaim.org/projects/million/) analysis found accessibility errors rising for the first time in six years, which WebAIM attributes in part to automated and AI-assisted coding. See [Training data](https://artificia11y.ds.house/inclusive-development/training-data/) for how that degraded web feeds back into the next generation of models. > [!SELF-ADVOCATE] > > Arguments that something is better for everyone are useful for getting buy-in, but they should not be the only reason a team acts. Access is worth building even when it helps only disabled users, because exclusion is the harm being prevented. Watch for the failure mode where a team builds the version of accessible AI that is most convenient for them, such as an alt-text button that is hard to find or a transcript that lags the audio, and then counts it done. The test is whether you can actually accomplish the task from start to finish, not whether a checkbox was ticked. Anna E. Cook makes the same point about [scores and dashboards](https://annaecook.com/writing/2026/1/5/designing-accessibility-for-real-use-not-dashboards), that a high accessibility score proves nothing if you cannot complete a core task like buying groceries or submitting a job application. Access is proven by people with disabilities using the product, not by a number. > > Be wary, too, of the opposite move, where accessibility is invoked for disabled people to wave through AI they did not ask for. When NaNoWriMo [defended generative AI](https://arstechnica.com/information-technology/2024/09/generative-ai-backlash-hits-annual-writing-event-prompting-resignations/) by calling opposition to it "classist and ableist," it spoke on behalf of disabled writers rather than with them. Disabled people's relationship to AI is genuinely mixed. The same technology that lets someone [build the assistive tool they always wanted](https://blakewatson.com/journal/i-used-claude-code-and-gsd-to-build-the-accessibility-tool-ive-always-wanted/) can also be used as an excuse to skip building real accessibility. Center agency, and let people decide for themselves. ## Where to go next - [What is AI accessibility?](https://artificia11y.ds.house/foundations/what-is-ai-accessibility/) covers the scope and the two surfaces. - [The feedback loop](https://artificia11y.ds.house/foundations/feedback-loop/) explains why this is continuous rather than one and done. - [WCAG applicability](https://artificia11y.ds.house/guidelines/wcag-applicability/) covers what the standards do and do not reach. ## Further reading - Harvard Gazette on [why AI fairness conversations must include disabled people](https://news.harvard.edu/gazette/story/2024/04/why-ai-fairness-conversations-must-include-disabled-people/). - Glazko and colleagues, ["An Autoethnographic Case Study of Generative AI's Utility for Accessibility"](https://dl.acm.org/doi/abs/10.1145/3597638.3614548) from ASSETS 2023. - Mason Marks, ["Algorithmic Disability Discrimination"](https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3338209) from 2019. - The European Disability Forum's [Empowered by AI](https://www.edf-feph.org/projects/empowered-by-ai/) project.