EU AI Act Article 50 for developers: what to ship before 2 December 2026
Last verified: 20 October 2026. Technical guidance, not legal advice.
Short answer: if your product chats with people in the EU, it must tell them they're dealing with AI. That has applied since 2 August 2026. If your product generates images, audio, video or text, that output must carry a machine-readable label showing it's AI-generated. AI systems already on the market before August have until 2 December 2026 for the labelling part. The rules apply to companies outside the EU too, as long as their systems serve people in the EU.
This guide turns the legal text into an engineering checklist.
Who has to do what
Article 50 of the AI Act (Regulation (EU) 2024/1689) has four duties. Two fall on providers (whoever builds or offers the AI system) and two on deployers (whoever uses it in their own product or publishing).
| # | Situation | Who | What you must ship | Exception |
|---|---|---|---|---|
| 50(1) | AI that interacts with people: chatbots, voice agents, AI agents | Provider | Tell people they're interacting with an AI system | When it's obvious to a reasonably well-informed person from the context |
| 50(2) | AI that generates synthetic audio, image, video or text | Provider | Mark outputs in a machine-readable format so they're detectable as AI-generated or manipulated | Assistive or standard editing that doesn't substantially alter the input |
| 50(3) | Emotion recognition or biometric categorisation | Deployer | Inform the people exposed to it | Some law-enforcement uses |
| 50(4) | Deepfakes, and AI text published to inform the public on matters of public interest | Deployer | Disclose that the content is AI-generated or manipulated | Text that a human reviewed and someone took editorial responsibility for |
The information must be given clearly and distinguishably, at the latest at the first interaction or exposure, and must meet accessibility requirements (Article 50(5)).
Most startups are a provider under 50(1) and 50(2) as soon as they ship a chatbot or a generation feature under their own name, even if the model comes from OpenAI, Anthropic or another vendor.
Key dates
| Date | What happened or happens |
|---|---|
| 10 June 2026 | Final voluntary Code of Practice on Transparency of AI-Generated Content published, including three EU icons for labelling AI content |
| 20 July 2026 | Commission adopted its Article 50 guidelines |
| 2 August 2026 | Article 50 obligations apply |
| 2 December 2026 | Machine-readable marking (50(2)) due for generative AI systems already on the market before 2 August 2026 |
Content you generated and published before 2 August 2026 doesn't need labelling retroactively.
Fines: up to €15 million or 3% of worldwide annual turnover. Signing the Code of Practice gives a degree of presumption of conformity; non-signatories must show compliance in other ways.
The engineering checklist
1. Disclosure for chatbots and agents (50(1))
- A visible, persistent label in the chat UI, for example "You're chatting with an AI assistant"
- The first message from the assistant says it's an AI
- Voice agents say it out loud at the start of the call
- Translated into your users' languages
- Accessible: real text, not only an icon; readable by screen readers
- No human-sounding persona that hides the AI. "Hi, I'm Sarah, Acme's AI assistant" is fine; "Hi, I'm Sarah" alone is not.
See the step-by-step tutorial: AI chatbot disclosure in React.
2. Machine-readable labels on generated content (50(2))
The law asks for machine-readable marking, not a specific technology. Guidance and the Code of Practice point to a layered approach, because each technique alone can be removed:
| Layer | What it is | Strength | Weakness |
|---|---|---|---|
| C2PA metadata ("Content Credentials") | A cryptographically signed manifest inside the file, stating it's AI-generated, by which tool, and when | Verifiable, an open standard, widely supported | Stripped by screenshots, re-encoding and many CDNs |
| Invisible watermark | A signal in the pixels or audio itself | Survives many edits | Vendor-specific; coverage varies by media type |
| Fingerprinting or logging | Hash of the output stored on your side | Proof you can show later | Not visible in the file itself |
Checklist:
- Images: C2PA manifest with an AI-generated assertion, plus an invisible watermark where available
- Audio and video: C2PA in the container. Watermark where available, since coverage is thinner.
- Text: no file-level standard exists. Use API response metadata or headers, a watermark where available (for example SynthID text), and a log of what you generated.
- Keep labels intact: check that your image optimisation, CDN and upload pipeline don't strip metadata
- Log every labelled output (hash, model, method, timestamp) so you can prove it later
3. Deepfakes and public-interest text (50(4)), if you publish
- Visible label on deepfake images, video and audio. The Code of Practice icons can serve this purpose.
- AI-written articles on public-interest topics are labelled, unless a named person or company took editorial responsibility after human review
4. Evidence
- A one-page record of which features are covered, how, and since when
- Ready to answer an enterprise customer's "Are you AI Act compliant?" question
Common mistakes
- "OpenAI or Google labels the images, so we're covered." Their labels cover their own output and can be stripped by your pipeline. Your chatbot disclosure and evidence are still on you.
- "We're not in the EU." The rules follow your users, not your office.
- "Text doesn't count." Generated text is synthetic content under 50(2).
- Adding C2PA, then serving images through a CDN that removes it. Test the final URL your users actually receive, not the file you generated.
- Disclosure hidden in the terms of service. It must be clear at the first interaction.
FAQ
Does Article 50 apply to companies outside the EU? Yes, when their AI system serves people in the EU.
What's the deadline for labelling AI-generated images? Article 50 applies from 2 August 2026. Generative systems already on the market before then have until 2 December 2026 for machine-readable marking.
Is a watermark required? The law requires machine-readable marking, not a specific technique. The Code of Practice describes a layered approach (metadata plus watermarking) because single methods can be removed.
Do I need to label AI text? Yes, under 50(2), if your system generates it. Deployers who publish AI text on public-interest matters must also disclose it, unless it went through human editorial review with someone taking responsibility.
What are the fines? Up to €15 million or 3% of worldwide annual turnover.
Sources
- Regulation (EU) 2024/1689 (AI Act), Article 50, on EUR-Lex
- Cooley: EU AI Act transparency obligations take effect 2 August 2026
- Freshfields: The final Code of Practice on transparency of AI-generated content
- ACT: What startups and SMEs should know as the AI Omnibus enters into force
Want this done in one npm install? Label50 handles disclosure, labelling and evidence for any AI stack. Join the waitlist →
This guide is technical information, not legal advice. Consult a lawyer about your specific obligations.