Blog Post
EU AI Act Labelling: What Your Product UI Must Show Now
By Emilia Veras. Last updated: August 13, 2026
The EU AI Act's transparency rules under Article 50 became applicable on August 2, 2026. If your product generates or publishes AI content for people in the EU, you now have to disclose it in the interface itself: a visible, plain-language label that a person can perceive without any special tools, shown on first exposure. A sparkle icon on its own does not satisfy this. Fines reach 15 million euros or 3% of worldwide annual turnover, whichever is higher.
The obligation lands in your UI. Not in a policy PDF, not in your terms page. In a component that has to render correctly on a 320px screen, survive dark mode, and be announced by a screen reader. That makes this a design QA problem wearing a compliance hat.
What Article 50 Actually Requires
Article 50 of the EU AI Act (Regulation (EU) 2024/1689) sets four separate duties. Under 50(1), systems that interact directly with people must make clear the person is talking to an AI (provider duty). Under 50(2), generative outputs must be marked in a machine-readable format and detectable as AI-generated (provider duty). Under 50(3), people exposed to emotion recognition or biometric categorisation must be informed (deployer duty). Under 50(4), deepfakes must be disclosed, as must AI-generated text published to inform the public on matters of public interest (deployer duty). Most product teams are deployers: you wrap an API and ship a feature, so your obligation is the part users can see.
| Provision | What it requires | Who it lands on |
|---|---|---|
| Article 50(1) | Systems that interact directly with people must make clear the person is talking to an AI | Provider |
| Article 50(2) | Generative outputs (text, image, audio, video) must be marked in a machine-readable format and detectable as AI-generated | Provider |
| Article 50(3) | People exposed to emotion recognition or biometric categorisation must be informed | Deployer |
| Article 50(4) | Deepfakes must be disclosed, and AI-generated text published to inform the public on matters of public interest must be disclosed | Deployer |
Machine-Readable Marking Is Not Enough
A watermark or C2PA metadata does not satisfy the EU AI Act on its own. Article 50(2) requires providers to mark generative outputs in a machine-readable format that is "effective, reliable, robust and interoperable," which is a separate duty from the visible disclosure a deployer owes under Article 50(4). The Commission's Article 50 FAQ states that deployers must disclose deepfakes in a clear and distinguishable manner that is understandable and perceivable by natural persons, for example with visible or audible labels, without need for any specific technical tools. Metadata a person cannot see is not a disclosure to that person. The FAQ also fixes the timing: disclosure must happen upon first exposure at the latest. That rules out a label that appears after an animation settles, a badge revealed on hover, and a disclosure that only exists on a detail view the user may never open.
What Counts as a Deepfake Under the AI Act
A deepfake under the EU AI Act is AI-generated or manipulated content that closely resembles a real or plausible person, object, or event and falsely appears authentic. Three cumulative criteria apply: a high level of similarity between the content and the subject it simulates, a simulated subject that exists or could plausibly exist, and content that falsely appears to a person to be authentic or truthful. Context is part of the test, so obvious visual effects in a film may fall outside the definition. Artistic, satirical, and fictional works get a lighter obligation. For the text duty in 50(4)(b), content is exempt where it went through human review or editorial control: a deliberate examination of the substance by a qualified person, plus the authority to approve, alter, or reject it. Spell-checking does not qualify.
The Sparkle Icon Does Not Count
A sparkle icon on its own is not a valid AI disclosure under Article 50, because the rule requires a signal a person can actually understand. Nielsen Norman Group tested the sparkle icon in isolation with 107 participants. Interpretations scattered: 16.82% read it as save or favorite, 16.82% as visual effects, 11.22% as new or important information. Zero participants mentioned AI or artificial intelligence (Kate Kaplan, NN/g, September 2024). The Commission has published an official EU icon set for labelling AI-generated content and states that using an icon does not establish legal compliance by itself. The practical rule is icon plus text: "AI-generated" next to a mark is legible to a first-time user, a screen reader, and a regulator.
What a Compliant Label Looks Like in the UI
Read as an interface spec, a compliant disclosure has to be perceivable without tools (rendered on screen or audible, not metadata and not a tooltip only), written in plain language, present on first render, clear and distinguishable with adequate contrast and size, exposed to assistive technology with a real accessible name, and persistent when the content is reshared or downloaded. That last one is the row most teams have not designed for: a label painted into your page layout does not travel with a downloaded asset. IBM's Carbon Design System AI Label is a production component with documented placement rules across form fields, tables, and dashboard tiles.
| Requirement | What it means in your component |
|---|---|
| Perceivable without tools | Rendered on screen or audible. Not metadata, not a tooltip only, not a title attribute |
| Plain language | Literal wording such as "AI-generated." Not a symbol alone, not a brand name for your model |
| On first exposure | Present on initial render, before or with the content. Not after a delay, not on hover, not one level deep |
| Clear and distinguishable | Adequate contrast and size against its own background. Not 10px grey on grey in a footer |
| Accessible | Exposed to assistive technology with a real accessible name, not a decorative icon with empty alt |
| Persistent | Survives resharing and download. A label that lives only in your page chrome is gone the moment the asset leaves |
The Label Is a Component, and Components Break
An AI disclosure that is designed correctly can still ship non-compliant, because the label is a small, text-bearing, contrast-sensitive component added to every surface that renders generated content, and that is precisely the class of element that breaks between design and production. The failure modes are ordinary and every one turns a compliant design into a non-compliant build: contrast drift when a badge designed on white ships on a grey card or a gradient, dark mode tokens applied unconditionally, overflow clipping, truncation that turns "AI-generated" into "AI-gener...", responsive reflow that pushes an absolutely positioned badge off-canvas on mobile, decorative markup with no accessible name, late render after hydration so the label is not present on first exposure, and localisation that breaks the layout. These are the standard catalogue of visual bugs, and they pass functional tests without complaint.
Deadlines and Penalties
The EU AI Act transparency obligations became applicable on August 2, 2026, and generative systems already on the market before that date have until December 2, 2026 to meet the machine-readable marking requirement. The Commission published the final Code of Practice on marking and labelling AI-generated content in June 2026 and concluded in July 2026 that it is adequate for demonstrating compliance with Articles 50(2), (4) and (5). Article 50 became applicable on August 2, 2026 and was not deferred by the Digital Omnibus. Generative systems already on the market before that date have until December 2, 2026 for the machine-readable marking requirement. Under Article 99, non-compliance carries fines of up to 15 million euros or 3% of total worldwide annual turnover, whichever is higher, with the lower of the two applying to SMEs and startups. Adhering to the Code of Practice is currently the only EU-wide tool assessed as adequate, regardless of where you are established.
| Date | What happens |
|---|---|
| June 10, 2026 | Commission publishes the final Code of Practice on marking and labelling AI-generated content |
| July 2026 | Commission and AI Board conclude the Code is adequate for demonstrating compliance with Articles 50(2), (4) and (5) |
| August 2, 2026 | Article 50 transparency obligations become applicable |
| December 2, 2026 | Grace period ends for machine-readable marking on generative systems already on the market before August 2, 2026 |
This Is Not Only an EU Problem
Like the European Accessibility Act, the AI Act follows the user rather than the company. Four other jurisdictions landed comparable rules inside eighteen months: China (September 1, 2025) requires explicit visible labels plus implicit labels such as watermarks; South Korea's AI Basic Act (January 22, 2026) requires advance notice that AI is in use plus disclosure for outputs hard to distinguish from reality; India's IT Rules amendments (February 2026) require labelling and provenance for synthetic media; and California's AI Transparency Act, SB 942 as amended by AB 853 (August 2, 2026), requires manifest and latent disclosures for providers above one million monthly users. The interface requirement converges even where the wording differs.
| Jurisdiction | In force | Shape of the rule |
|---|---|---|
| China | September 1, 2025 | Explicit visible labels plus implicit labels such as watermarks on AI text, image, audio, video, and virtual scenes |
| South Korea | January 22, 2026 | AI Basic Act: advance notice that AI is in use, plus disclosure for outputs hard to distinguish from reality |
| India | February 2026 | IT Rules amendments: labelling and provenance for synthetic media, with a proposal to keep the label visible for the full duration of the content |
| California | August 2, 2026 | AI Transparency Act (SB 942, amended by AB 853): manifest and latent disclosures, detection tooling, for providers above one million monthly users |
How to Verify Your Label Actually Ships
Verifying an AI label means checking it on the rendered page at every breakpoint, in both themes, on every surface that emits generated content, and confirming assistive technology announces it. Inventory every surface that renders AI output, including the exports teams usually forget. Check contrast on the real background rather than the swatch with a contrast checker. Test the label at every breakpoint in both themes with a responsive tester, since absolute positioning, flex truncation, and overflow clipping only reveal themselves at specific widths. Confirm assistive technology announces it using a free accessibility checker and screen reader testing. Verify persistence by downloading the asset and opening it outside your product. Then re-check on every release, because a token change or a card refactor can silently break the label months after it passed.
Frequently Asked Questions
When did the EU AI Act labelling rules take effect?
The Article 50 transparency obligations became applicable on August 2, 2026, and were not postponed by the Digital Omnibus. Generative systems already on the market before that date have until December 2, 2026 to meet the machine-readable marking requirement. Content generated before August 2, 2026 does not need retroactive labelling.
Is a sparkle icon enough to label AI-generated content?
No. The disclosure has to be understandable and perceivable by a person without special tools, which means a plain-language label such as "AI-generated," not a symbol on its own. The European Commission publishes an official icon set and states that using an icon does not establish legal compliance by itself. Nielsen Norman Group tested the sparkle icon with 107 participants and none of them associated it with AI.
What are the fines for failing to label AI content in the EU?
Under Article 99 of the AI Act, breaching the Article 50 transparency obligations carries administrative fines of up to 15 million euros or 3% of total worldwide annual turnover for the preceding financial year, whichever is higher. For SMEs and startups the cap is whichever figure is lower. Enforcement is handled by national market surveillance authorities in each member state.
Does the EU AI Act apply to companies outside the EU?
Yes. Like the European Accessibility Act, the AI Act follows the user rather than the company. If your product generates or publishes AI content that reaches people in the EU, the Article 50 obligations apply regardless of where your company is incorporated or hosted. Adhering to the Commission's Code of Practice on marking and labelling AI-generated content is currently the only EU-wide compliance tool assessed as adequate, and it applies regardless of place of establishment.
Do I need to label AI-assisted writing?
Not usually. Assistive editing such as grammar correction, spell-check, formatting, and cropping does not trigger the marking obligation. The text disclosure in Article 50(4)(b) applies to AI-generated text published to inform the public on matters of public interest, and it is exempt where a qualified person substantively reviewed the content with the authority to approve, alter, or reject it, and someone holds editorial responsibility. A spell-check pass does not meet that bar.