AI and Accessibility Law: Accessible AI Notices, Labels and Chatbots

EU AI Act: Article 50(5) and the Accessible Notice Duty

Article 50 of Regulation (EU) 2024/1689 sets the AI Act's transparency duties. Providers must design AI systems that interact directly with people so those people are told they are dealing with an AI system, unless that is obvious (50(1)), and must mark synthetic audio, image, video and text in a machine-readable, detectable way (50(2)). Deployers must inform people exposed to emotion recognition or biometric categorisation (50(3)), disclose deep fakes, and disclose AI-generated text published to inform the public on matters of public interest unless it had human review or editorial control (50(4)).

Paragraph 5 governs how all four are delivered. The information must reach the people concerned in a clear and distinguishable manner at the latest at the first interaction or exposure, and then:

“The information shall conform to the applicable accessibility requirements.”

The duty therefore falls on providers for the interaction notice and content marking, and on deployers for emotion recognition, deep-fake and public-interest text disclosures. Recital 132 adds that the characteristics of people in vulnerable groups due to age or disability should be taken into account, and that notifications “should be provided in accessible formats for persons with disabilities”.

When it applies

Article 50 has applied since 2 August 2026 under Article 113, and the Commission's Guidelines (paragraph 153) confirm it applies from that date regardless of when a system was placed on the market. The Digital Omnibus on AI, Regulation (EU) 2026/1744, published on 24 July 2026 and in force from 27 July 2026, moved only one piece: providers of generative systems placed on the market before 2 August 2026 have until 2 December 2026 to comply with the Article 50(2) marking duty (new Article 111(4)). The interaction disclosure under 50(1) had to be in place from 2 August 2026. The Omnibus did not amend paragraph 5 at all; its only change to Article 50 replaces paragraph 7 on codes of practice.

What the Commission's Guidelines say about accessibility

The Guidelines then get specific about format. Paragraph 34 says the information must be provided in an accessible format for persons with disabilities, as Article 50(5) and recital 132 require, and a footnote notes that where a system interacts with the general public, vulnerable groups are included. Paragraph 35 adds that a provider who can foresee interaction with vulnerable groups should provide the information in an accessible format. Paragraph 37 recommends multimodal disclosure:

• Text labels such as “You are interacting with an AI system”, first-turn chatbot greetings and persistent badges placed near the input or output field.
• Spoken statements in voice contexts.
• Earcons, which may help visually impaired users recognise the AI but “are not considered sufficient by themselves”.
• Combinations of modes to ensure accessibility.

Paragraph 38 says a disclosure only in terms and conditions, a URL or documentation is not enough on its own, and paragraph 142 rules out information hidden under layers of menu options. Paragraph 143 sets the timing: at least once at the start of each interactive session, such as when launching a conversation with a chatbot, and per output for generated content.

The AAC exemption. Among the examples for the Article 50(2) exceptions, the Guidelines say content that only transforms authentic human input through assistive technologies that let persons with disabilities communicate, such as augmentative and alternative communication (AAC), does not need to be marked as AI-generated. Edits made to align content with accessibility requirements count as standard editing. A synthetic voice that speaks for a disabled person does not need an AI watermark.

Article 16(l): high-risk AI must itself be accessible

Article 16 lists what providers of high-risk AI systems must do. Point (l) reads: “ensure that the high-risk AI system complies with accessibility requirements in accordance with Directives (EU) 2016/2102 and (EU) 2019/882.” This is a duty on the system, not just its notices, and it falls on providers only: Article 26, which sets deployers' duties, has no accessibility wording. Recital 80 explains why, citing the UN Convention on the Rights of Persons with Disabilities and adding that providers should ensure compliance “by design”.

The Omnibus pushed the dates back because of the delayed availability of standards, common specifications and alternative guidance (recital 40). Chapter III, Sections 1 to 3, which contain Article 16, now apply from 2 December 2027 for high-risk systems under Article 6(2) and Annex III, and from 2 August 2028 for systems under Article 6(1) and Annex I. The original dates were 2 August 2026 and 2 August 2027. High-risk systems already on the market before those dates are caught only after significant changes in their design, but providers and deployers that are public authorities must comply by 2 August 2030 (new Article 111(2)).

No harmonised AI standard deals with accessibility. The Commission's 2023 standardisation request for the AI Act, C(2023) 3215, mentions disability once, asking that the UN Convention be taken into account (recital 14). For 16(l) and 50(5), EN 301 549, the harmonised standard for the EAA and the Web Accessibility Directive, remains the practical measuring stick.

The Code of Practice names EN 301 549 and WCAG 2.1 AA

The Code of Practice on Transparency of AI-generated Content, final text of 10 June 2026, is voluntary, and adherence “does not constitute conclusive evidence of compliance”. But on 8 July 2026 the Commission gave its opinion that the Code adequately covers the obligations in Articles 50(2), (4) and (5), and the AI Board adopted its assessment the next day. It is the most concrete statement of what an accessible AI label looks like:

• Providers, Measure 2.3.Interfaces that show detection results must be accessible to persons with disabilities under the EAA and the Web Accessibility Directive. Signatories are encouraged to implement relevant standards, “including but not limited to” EN 301 549 and WCAG 2.1 Level AA.
• Deployers, objective (c). Disclosure should be easily perceivable and understandable by everyone, including persons with disabilities, children and the elderly.
• Deployers, Measure 1.1. Signatories will ensure accessible disclosure, including audio descriptions or alternative cues for visual disclosure elements; tactile and haptic cues for audio-only content, such as a vibration before deep-fake audio; high contrast icons or equivalent labels and screen reader compatibility, including for users with colour vision deficiencies; and detectability of the icon or label by assistive technologies. EN 301 549 or WCAG 2.1 are encouraged where the service falls within their scope.

Penalties

Article 99(4) sets fines of up to EUR 15 million or 3% of total worldwide annual turnover, whichever is higher, and expressly lists providers' obligations under Article 16 and the transparency obligations of providers and deployers under Article 50. For SMEs, and since the Omnibus for small mid-caps, the lower of the two applies. EU institutions face up to EUR 750,000. National market surveillance authorities enforce, the AI Office handles systems built on its own general-purpose models, and the European Data Protection Supervisor covers EU institutions.

Other accessibility and disability provisions in the AI Act

Article 5(1)(b)

Prohibits AI that exploits vulnerabilities due to age, disability or a specific social or economic situation to materially distort behaviour. Applies since 2 February 2025; fines up to EUR 35 million or 7% under Article 99(3).

Article 60(4)(g)

Real-world testing must appropriately protect subjects who belong to vulnerable groups due to age or disability.

Article 71(6)

The EU database of high-risk systems "shall comply with the applicable accessibility requirements".

Article 95(2)(e)

Voluntary codes of conduct may cover the impact of AI on vulnerable persons, including accessibility for persons with a disability.

Recital 80

Cites the UN Convention on the Rights of Persons with Disabilities and says providers should ensure compliance with the accessibility directives "by design".

Recital 132

Transparency notices "should be provided in accessible formats for persons with disabilities".

Recitals 142 and 165

Encourage funding for AI that increases accessibility, and voluntary codes on accessibility to persons with disability.

Articles 13(2) and 77(1) also use the word “accessible”, but in the sense of available or understandable, not disability access.

Colorado: Reasonably Accessible AI Notices and a WCAG 2.2 Draft Rule

The statute.SB26-189, signed on 14 May 2026, repealed and re-enacted Colorado's AI framework as an automated decision-making technology law that takes effect on 1 January 2027. Section 6-1-1704 requires a clear and conspicuous pre-use notice before covered ADMT materially influences a consequential decision, a public notice reasonably accessible at points of consumer interaction, and an adverse-outcome disclosure within 30 days. Subsection (8) adds the accessibility duty: a deployer or developer must provide these notices and disclosures in a manner “reasonably accessible to consumers with disabilities and consumers with limited English proficiency”, consistent with applicable state and federal law.

The chatbot law. HB26-1263, approved on 29 May 2026, requires an operator of a conversational AI service to disclose clearly and conspicuously that the service is artificial intelligence: at the first interaction each day, every three hours or as a persistent disclosure, and whenever a user asks whether it is human. Duties apply from 1 January 2027. For minors, screen products need a persistent visible disclaimer and screenless products an intermittent audio one. The statute itself has no accessibility wording. The interim draft rules renumber the chatbot section as 6-1-1710, so check the codified number before citing it.

The draft rule.The Attorney General's draft rules, 4 CCR 904-6, apply Rule 3.2 to notices from deployers to consumers and from chatbot operators to users. Rule 3.2(A)(2) as proposed on 11 August 2026 requires disclosures to be reasonably accessible to consumers or users with disabilities, including through digital accessibility tools. For information online, the deployer or operator must follow generally recognised industry standards, such as WCAG 2.2 of 12 December 2024, incorporated by reference in Rule 14.1 without later amendments. In other contexts, it must explain how a person with a disability can get the disclosure in an alternative format. The 6 October 2026 interim draft changed the wording to “such as, but not limited to”. Neither draft names a conformance level. Rule 3.2(A) also asks for plain language, the languages ordinarily used, and a format readable on all devices including smaller screens. The wording mirrors California's § 7003(b)(3), with WCAG 2.2 in place of 2.1.

Enforcement. Under C.R.S. 6-1-1706 the Attorney General enforces through the Colorado Consumer Protection Act, and the disclosure duties are enforceable exclusively by the Attorney General. A violation is a deceptive trade practice. Before acting, the Attorney General gives notice and a 60-day cure period where cure is possible, but not for knowing or repeated violations. There is no private right of action.

We walk through the draft rule's wording, the “such as, but not limited to” change and what it means for chatbot teams in Colorado's AI rules and WCAG 2.2 accessible disclosures.

California: ADMT Notices Under WCAG 2.1 and Accommodation in AI Hiring

CCPA regulations.The California Privacy Protection Agency's regulations on cybersecurity audits, risk assessments and automated decision-making technology were approved by the Office of Administrative Law on 22 September 2025, effective 1 January 2026. Section 7003(b)(3) of title 11 requires disclosures to be reasonably accessible to consumers with disabilities; for notices online, the business must follow generally recognised industry standards, such as WCAG 2.1 of 5 June 2018, and in other contexts must explain how a consumer with a disability can get the information in an alternative format. Section 7220(b)(1) applies that rule to the ADMT Pre-use Notice, and § 7221 applies it to communications about appeals of ADMT decisions. A business already using ADMT for a significant decision must comply no later than 1 January 2027 (§ 7200(b)).

Administrative fines under Civil Code 1798.155(a), as adjusted for 2025, are up to $2,663 per violation or $7,988 per intentional violation or violation involving minors under 16. Our California accessibility law guide covers the Unruh Act exposure that sits alongside these rules.

Civil Rights Council employment rules.The Council's automated-decision system regulations took effect on 1 October 2025. 2 CCR § 11016(c)(5) says an ADS that measures an applicant's skill, dexterity, reaction time or other abilities may discriminate against people with certain disabilities, and § 11016(d)(1) says the same of an ADS that analyses tone of voice, facial expressions or other physical characteristics. Both add that an employer may need to provide reasonable accommodation to avoid unlawful discrimination. The Council's Final Statement of Reasons says the existing requirement to provide reasonable accommodations applies to the use of automated-decision systems. The Civil Rights Department adds that ADS assessments, including puzzle games, that elicit information about a disability may be an unlawful medical inquiry, and that ADS data must be kept for four years.

SB 243 has no accessibility clause. California's companion chatbot law requires a clear and conspicuous notification that a companion chatbot is artificially generated and not human, where a reasonable person could be misled, and gives a private right of action with damages of at least $1,000 per violation. Business and Professions Code sections 22601 to 22605 contain no accessibility or disability wording, and customer-service bots are excluded. Making that notice accessible is good practice, not an SB 243 requirement.

Other United States Rules

New York City.Local Law 144 of 2021 regulates automated employment decision tools. The Department of Consumer and Worker Protection's rule, in force and enforced since 5 July 2023, says the candidate notice “must include instructions for how an individual can request an alternative selection process or a reasonable accommodation under other laws, if available” (6 RCNY § 5-304(a)). The rule does not require an alternative process itself, and the notice is due ten business days before use.

Federal ADA guidance.The Department of Justice's 2022 guidance on algorithms, AI and disability discrimination in hiring is still published on ada.gov. It says employers must provide requested reasonable accommodations that let applicants and employees with disabilities meet the tool's standards, unless doing so would be an undue hardship. The EEOC's companion document from May 2022 now returns a 404. Removing guidance does not change the ADA: Title I still governs AI in employment, and our AI accessibility lawsuits guide tracks how claims are being brought. State and local government chatbots are web content under the ADA Title II web rule, which uses WCAG 2.1 AA, and federal agencies' AI interfaces fall under Section 508, which has no AI-specific guidance.

Texas. The Texas Responsible Artificial Intelligence Governance Act (HB 149), in effect since 1 January 2026, requires governmental agencies to disclose AI interaction in plain language, clearly and conspicuously and without dark patterns, but says nothing about accessibility. Its disability hook is a ban on developing or deploying AI with the intent to unlawfully discriminate against a protected class, which includes disability; disparate impact alone is not enough.

Illinois (not confirmed from the statute). Secondary sources report that HB 3773 amended the Illinois Human Rights Act from 1 January 2026 to bar AI use that has a discriminatory effect on protected classes, which include disability, and that draft notice rules referring to accessibility and accommodation were withdrawn in June 2026. We could not read the statute from the primary source, so it is not in the table.

Canada and the United Kingdom

CAN-ASC-6.2:2025.Accessibility Standards Canada published “Accessible and Equitable Artificial Intelligence Systems” in December 2025. Its overview says that if AI affects workers, customers or the public, it must meet accessibility standards, and that organisations should use accessible formats when telling the public about their use of AI. It is reported to be voluntary unless adopted in regulation. For federal institutions, the Treasury Board Directive on Automated Decision-Making already requires algorithmic impact assessment results to be published “in an accessible format” (6.1.1). More on the federal regime is on our Canada accessibility law page.

United Kingdom. The UK has no AI Act. The Equality Act 2010 duty to make reasonable adjustments applies to AI services as to any other, and public sector chatbots on websites and apps fall under the 2018 public sector accessibility regulations. See the UK accessibility law guide.

What To Do: An Accessibility Checklist for AI Notices, Labels and Chat

The laws name different WCAG versions, but they converge on the same design. Build the notice, the label and the chat interface to WCAG 2.2 Level AA and EN 301 549, and you cover the technical requirements of the EU regime where it applies, Colorado's draft and California's rule. The steps below come from the Commission's Guidelines, the Code of Practice and the US texts.

1. Disclose in context, at the first interaction. Show the AI notice when the session starts, near the input or output field, and keep a persistent label. A line in the terms and conditions is not enough.
2. Make the label real text. If you use an AI icon, give it a text alternative that states its meaning, keep it in the accessibility tree, and never rely on colour alone. Aim for 4.5:1 contrast for text and 3:1 for the icon.
3. Use more than one sense. Pair visual labels with an audio or text cue, give voice assistants a spoken statement, and give audio-only content a visual or haptic cue. An earcon can support recognition but cannot be the only signal.
4. Announce without stealing focus.Use a status message for the AI greeting, “responding” and “response complete”, so screen reader users hear them without losing their place.
5. Keep the chat operable and unobscured. Every control works from the keyboard, focus moves into the chat and back, and the floating launcher or AI banner never hides the focused element.
6. Fit every screen. Notices must reflow at 320 CSS pixels; Colorado and California both require readability on smaller screens.
7. Publish an alternative format route. Both US rules require information, in contexts other than online, on how a person with a disability can get the notice in an alternative format. Name a contact and a format list.
8. Offer an accommodation route around automated decisions. In hiring, tell candidates how to request an alternative process or reasonable accommodation before the tool is used, as New York City requires and California and the ADA expect.
9. Do not watermark assistive speech.AAC output that transforms a disabled person's own input falls outside the Article 50(2) marking duty.
10. Test with assistive technology. Check the label and the chat with a screen reader, keyboard only, high contrast and zoom. Our accessible AI chat guide maps each pattern to WCAG, and the WCAG 2.2 checklist covers the full set.

WCAG 2.2 criteria that matter most for AI notices and chat

EN 301 549 clauses beyond WCAG

Where the EAA or the Web Accessibility Directive applies, EN 301 549 reaches past web content. For AI products, the clauses to check in V3.2.1 are:

• 9.4.1.3 and 11.4.1.3 Status messages, on the web and in software.
• 5.1.3.1 Audio output of visual information, for kiosks and other closed functionality.
• 6.1, 6.2.1.1 and 6.2.1.2 Audio bandwidth for speech, real-time text, and concurrent voice and text, for voice bots and two-way voice.
• 11.5.2.3 and 11.7 Use of accessibility services and user preferences, for native apps.
• 12.1.1, 12.1.2, 12.2.2, 12.2.3 and 12.2.4 Documentation and support services, including effective communication, which matters wherever an AI assistant replaces human support.

Sources

Every provision, date and quotation on this page was read in the documents below, checked on 11 November 2026. Where only secondary reporting was available, the text says so.

1. Regulation (EU) 2024/1689 (Artificial Intelligence Act), OJ L, 12.7.2024: Articles 5, 16, 50, 71, 95, 99, 113 and recitals 80, 132, 142 and 165
2. Regulation (EU) 2026/1744 (Digital Omnibus on AI) of 8 July 2026, OJ L, 24.7.2026: new Articles 111(2), 111(4) and 113(c)
3. European Commission, Guidelines on the implementation of the transparency obligations under Article 50 of the AI Act, C(2026) 7702 final, 5 November 2026 (PDF)
4. European Commission, Article 50 Guidelines library page (published 20 July 2026, last updated 5 November 2026)
5. Code of Practice on Transparency of AI-generated Content, final text of 10 June 2026 (PDF)
6. European Commission, Code of Practice on AI-generated content policy page
7. European Commission, Opinion of 8 July 2026 on the assessment of the Code of Practice on transparency of AI-generated content
8. Commission Implementing Decision C(2023) 3215 final of 22 May 2023, the AI standardisation request (recital 14)
9. Colorado SB26-189, Automated Decision-Making Technology, signed act (C.R.S. 6-1-1704 and 6-1-1706)
10. Colorado General Assembly, SB26-189 bill page (signed 14 May 2026)
11. Colorado HB26-1263, Conversational Artificial Intelligence Services, Session Laws 2026 chapter 208 (approved 29 May 2026)
12. Colorado Attorney General, ADMT and Chatbot Act rulemaking page, 4 CCR 904-6 (checked 11 November 2026)
13. Colorado Attorney General, proposed rules filed 11 August 2026, Rules 1.3, 3.2 and 14.1 (Word document)
14. Colorado Attorney General, interim draft redline of 6 October 2026 (PDF)
15. W3C, Web Content Accessibility Guidelines (WCAG) 2.2, Recommendation of 12 December 2024
16. California Privacy Protection Agency, approved regulation text on CCPA updates, cybersecurity audits, risk assessments and ADMT: 11 CCR §§ 7003, 7200, 7220 and 7221 (PDF)
17. California Privacy Protection Agency, CCPA updates rulemaking status page (approved 22 September 2025, effective 1 January 2026)
18. California Privacy Protection Agency, 2025 CPI adjustment of administrative fine amounts
19. California Civil Rights Department, release of 30 June 2025 on the automated-decision system employment regulations
20. California Civil Rights Council, final text of the automated-decision system employment regulations, 2 CCR § 11016 (PDF)
21. California Civil Rights Council, Final Statement of Reasons for the automated-decision system regulations (PDF)
22. US Department of Justice, Algorithms, Artificial Intelligence, and Disability Discrimination in Hiring, 12 May 2022
23. New York City DCWP, notice of adoption of the rule on automated employment decision tools, 6 RCNY § 5-304 (PDF)
24. New York City DCWP, frequently asked questions on automated employment decision tools
25. Texas HB 149 (89th Legislature), Texas Responsible Artificial Intelligence Governance Act, enrolled text
26. Treasury Board of Canada Secretariat, Directive on Automated Decision-Making (section 6.1.1)
27. Accessibility Standards Canada, overview of CAN-ASC-6.2:2025, Accessible and Equitable Artificial Intelligence Systems