Inkluso banner on a dark green background quoting the second sentence of Article 50(5): “The information shall conform to the applicable accessibility requirements.”
The sentence most Article 50 guides stop just short of. Regulation (EU) 2024/1689, Article 50(5), second sentence.

Since 2 August 2026 the transparency duties in Article 50 of the EU AI Act have applied across the Union. Most of the commentary since then has been about who owes the disclosure and what counts as a deepfake. Less of it has been about the form the disclosure has to take, which is what Article 50(5) governs.

The clause is two sentences long. Most guides quote the first one.

What Article 50(5) says

From the Commission's own AI Act Service Desk:

"The information referred to in paragraphs 1 to 4 shall be provided to the natural persons concerned in a clear and distinguishable manner at the latest at the time of the first interaction or exposure. The information shall conform to the applicable accessibility requirements."

Because it refers to paragraphs 1 to 4, the second sentence reaches every disclosure in the article: the chatbot notice under 50(1), the machine-readable marking under 50(2), the emotion recognition and biometric categorisation notice under 50(3), the deepfake and public-interest text label under 50(4).

And because Article 50 applies irrespective of risk tier, this is the one place an accessibility obligation attaches to systems that are not high-risk at all. Article 16(l) already put accessibility on providers of high-risk systems. Article 50(5) extends it to everyone who owes a transparency duty.

If you are still working out which duties landed on 2 August and which the omnibus moved, our guide to what actually applies from 2 August sets out the timeline.

Which accessibility requirements are "applicable"?

The Act does not define the term inside Article 50, but it names the instruments elsewhere. Article 16(l) requires providers of high-risk systems to "ensure that the high-risk AI system complies with accessibility requirements in accordance with Directives (EU) 2016/2102 and (EU) 2019/882". Recital 80 repeats the pair and adds that providers "should ensure compliance with these requirements by design".

Those two are the Web Accessibility Directive, which covers public sector bodies, and the European Accessibility Act, which has applied to services since 28 June 2025 and is transposed nationally: the BFSG in Germany, the BaFG in Austria, Act 424/2023 Sb. in Czechia.

Recital 132 addresses Article 50 directly. Notifications under these transparency duties "should be provided in accessible formats for persons with disabilities", and when implementing the obligation, "the characteristics of natural persons belonging to vulnerable groups due to their age or disability should be taken into account, to the extent the AI system is intended to interact with those groups as well". A public-facing chatbot is aimed at the general public, and that audience includes screen reader users.

So Article 50(5) does not create a new accessibility standard. It points at whichever one already binds you, and pulls your AI disclosure into that scope.

Who this binds, and who it does not

Public sector bodies are covered by the Web Accessibility Directive, and the disclosure inherits it. Providers of services in the European Accessibility Act's scope, which includes e-commerce, consumer banking, e-books, electronic communications and elements of passenger transport services, are covered by the EAA, and the disclosure inherits that.

The EAA's micro-enterprise carve-out covers service providers with fewer than ten staff and annual turnover or balance sheet total of two million euro or less. Sellers of the products in scope get no equivalent exemption.

If no accessibility regime binds you, the second sentence has nothing to attach to and you are left with the first. That is narrower, though the next section is a reason not to lean on it.

The obviousness carve-out has a gap in it

Article 50(1) does not require a disclosure where it is already obvious that a person is dealing with AI. Obviousness is measured against a person who is "reasonably well-informed, observant and circumspect", in the context of use.

Recital 132 then requires that the characteristics of people in vulnerable groups due to age or disability be taken into account when implementing that obligation. A visual signal that makes a chatbot obvious to a sighted user, an avatar, a colour treatment, a stylised interface, is not obvious to a screen reader user. The carve-out may not hold for part of the audience.

The Commission's final guidelines point the same way, treating the requirements as stricter where children or other vulnerable groups are in the intended audience.

What the Commission's guidelines already rule out

The Commission's draft guidelines of 8 May 2026, on Greenberg Traurig's reading, rejected three patterns outright, and the final guidelines adopted on 20 July 2026 hold that line: a reference in the terms and conditions or product documentation, which "is not sufficient"; technical labels alone, since "metadata or watermarks alone do not meet the requirement either, users typically do not notice them at the point of interaction"; and vague terms such as "assistant" or technical descriptions like "this system uses LLMs".

What they recommend instead is a combination of formats, including plain-language notices, audio cues, and persistent visual indicators, delivered at the point of interaction.

That recommendation is worth sitting with. The guidelines are already pointing past a purely visual disclosure, before you reach the accessibility sentence at all.

Six patterns that fail

Disclosure baked into the image as pixels. An "AI generated" mark rendered into the artwork carries no text. A screen reader announces nothing unless the alt text repeats it.

A badge or icon with no accessible name. An AI glyph next to a message is invisible to assistive technology without a programmatic label.

Low contrast. Grey caption text on a light background, or a caption over a busy photograph, can be present and unreadable. The 4.5:1 ratio for normal text in WCAG is the benchmark the national regimes work from.

Hover-only tooltips. A disclosure that appears on mouse hover is not reachable by keyboard and does not exist on touch. It usually also fails the timing test, arriving after first interaction rather than at it.

Toasts that vanish. A banner that auto-dismisses, in a container with no live region, is never announced. For that user it did not arrive at first exposure.

Terms, footers and settings pages. Already ruled out for everyone by the guidelines above. If the accessible fallback is "it is in the terms", there are two problems rather than one.

What a compliant disclosure looks like

It exists as text rather than as pixels. It is programmatically associated with what it describes, so the relationship survives non-visual navigation. Any badge or icon carrying it has an accessible name.

It meets the contrast the applicable national regime works from. It is reachable by keyboard and by touch, not only by pointer. It is announced at first interaction or exposure rather than after, which in a dynamic interface usually means a live region rather than an element that renders somewhere off the reading path.

For published media, the visible caption carries the weight, since the guidelines are explicit that machine-readable marking alone does not discharge the obligation.

There is no safe harbour, so the deliverable is evidence

No harmonised standard has been cited in the Official Journal under the European Accessibility Act, so no presumption of conformity is available. EN 301 549 V3.2.1, which maps to WCAG 2.1 level A and AA, is the benchmark in practice, and V4.1.0, which moves to WCAG 2.2, is a final draft rather than a version in force. Neither is a legal shield. The position differs under the Web Accessibility Directive, where Decision 2018/2048 does cite EN 301 549.

You cannot point at a standard and stop. What you can produce is a dated record that you tested. Worth knowing while planning: an automated scan reliably catches roughly 30 to 40 percent of WCAG issues. Focus order, meaningful sequence and complex widgets need a keyboard and a screen reader in the hands of a person.

What to do

  1. List every surface that owes a disclosure. Chat and voice interfaces, anything generating synthetic media, anything doing emotion recognition or biometric categorisation, any AI-drafted text published to inform the public.
  2. Record, per surface, whether you are the provider or the deployer. Article 50 assigns duties on that line, and most inventories do not have the column.
  3. Test each disclosure with a keyboard and a screen reader. Tab to it. Listen for it. Note whether it arrives before or after first interaction.
  4. Fix the pixels-and-icons cases first. Cheapest to fix, most common to fail.
  5. Record the date you tested and what you found. Under the AI Act, breaches of the transparency duties carry maximums of fifteen million euro or three percent of worldwide annual turnover, whichever is higher; for SMEs the cap is whichever of the two is lower. Under the national accessibility regimes there is a separate exposure with its own statutory maximums. The same artefact answers both, and it is a dated record rather than a claim.

For step 3, the automated layer is the cheap half and you can start it now: our free website accessibility checker tests a page against WCAG 2.1 AA, including the contrast and accessible-name failures that sink most disclosure badges.

Why these two regimes meet here

Article 16(l) had already made accessibility a design-time requirement for high-risk systems. What Article 50(5) adds is reach: a transparency duty that lands on systems of any risk tier, resolving into an accessibility requirement that many of the teams shipping those disclosures have never had to read.

Disclosure of interest: we work on both sides. Klariq covers the EU AI Act, Inkluso covers accessibility, which is how this clause reached our desk, and probably why it gets missed. Most teams have one of those two competences in the building. This sentence needs both.

Sources

See where your site stands

A free scan checks your site against WCAG 2.1 AA and shows the first findings in minutes.

Scan your site free Download a sample report

Related article: EU AI Act on 2 August 2026: exactly what changes

Not legal advice. Orientation, with the sources named so you can check them against your own facts.