What Article 50 of the EU AI Act is — the four disclosure rules behind ChatGPT watermarks
Article 50 of the EU AI Act is the transparency article, and it has applied since August 2, 2026. It sets four duties. AI systems that talk with people must tell them they are dealing with AI. Providers of generative AI must mark outputs in a machine-readable format so they can be detected as AI-generated. Deployers of emotion recognition or biometric categorisation must inform the people exposed. Deployers of deepfakes, or of AI-generated text published to inform the public on matters of public interest, must disclose that it is artificial. Breaches can cost up to €15 million or 3% of worldwide annual turnover, whichever is higher. It is the rule behind the ChatGPT text watermark OpenAI announced for the EU in October 2026
The three lines
- What — the EU AI Act transparency article; applied August 2, 2026; legacy systems reportedly have until December 2
- Four duties — disclose chatbots; machine-readable marks on generated content; disclose emotion recognition; label deepfakes
- Teeth — up to €15M or 3% of global turnover; South Korea's AI Basic Act has its own labeling rule
Key questions
- What does Article 50 of the AI Act require
- **Four duties split between providers and deployers.** | Paragraph | Who | What | |---|---|---| | 50(1) | Provider | Tell people they are interacting with AI | | 50(2) | Generative AI provider | Machine-readable, detectable marks on text, image, audio, video | | 50(3) | Deployer | Inform people exposed to emotion recognition or biometric categorisation | | 50(4) | Deployer | Disclose deepfakes and AI text published on public-interest matters |
- When does AI Act Article 50 apply
- **From August 2, 2026.** | Milestone | Date | |---|---| | Act enters into force | Aug 1, 2024 | | Prohibited practices | Feb 2, 2025 | | General-purpose AI duties | Aug 2, 2025 | | **Article 50** | **Aug 2, 2026** | | Legacy systems (reported) | Dec 2, 2026 |
- AI Act transparency fines
- **Up to €15 million or 3% of worldwide turnover, whichever is higher.** | Breach | Ceiling | |---|---| | Prohibited AI | €35M or 7% | | Article 50 and other duties | €15M or 3% | | Misleading information to authorities | €7.5M or 1% | For SMEs and startups the lower of the two applies.
Europe already has a law requiring AI-made text, images and voices to be identifiable as AI-made. It is Article 50 of the EU Artificial Intelligence Act, which began applying on August 2, 2026 — and it is why OpenAI said on October 5 that it will watermark ChatGPT text for EU users. The article is short, but it splits the duty precisely between those who build AI and those who use it.
1. Where Article 50 sits
The AI Act sorts AI into four risk tiers. Article 50 covers the transparency tier: systems that are not dangerous in themselves but cause harm if people do not know they are dealing with AI.
| Tier | Examples | Approach |
|---|---|---|
| Prohibited | Social scoring, untargeted face scraping | Banned (from Feb 2025) |
| High-risk | Hiring, credit scoring, medical AI | Conformity assessment |
| Transparency (Art. 50) | Chatbots, generative AI, deepfakes | Disclosure and marking |
| Minimal | Spam filters, game AI | Voluntary |
2. Four duties
| Para. | Who | Duty | Example |
|---|---|---|---|
| 50(1) | Provider | Tell people they are interacting with AI (unless obvious) | Customer-service bot |
| 50(2) | Generative AI provider | Mark outputs machine-readably so they are detectable as AI | Text watermark, image metadata |
| 50(3) | Deployer | Inform people subject to emotion recognition or biometric categorisation | In-store emotion cameras |
| 50(4) | Deployer | Disclose deepfakes; disclose AI text published on public-interest matters unless a human holds editorial responsibility | Synthetic ad video, AI-written news |
Paragraph 2 is the technical one. It requires marks that other systems can detect automatically, so a visible label alone is not enough — hence OpenAI's statistical text watermark. The editorial-responsibility exemption in paragraph 4 is aimed at newsrooms that use AI as an aid under human review. Clearly artistic, satirical or fictional deepfakes face a lighter duty: disclose in a way that does not spoil the work.
3. Timeline and penalties
| Milestone | Date |
|---|---|
| Act in force | Aug 1, 2024 |
| Prohibitions apply | Feb 2, 2025 |
| General-purpose AI model duties | Aug 2, 2025 |
| Article 50 applies | Aug 2, 2026 |
| Systems already on market (reported) | Dec 2, 2026 |
| Breach | Maximum fine |
|---|---|
| Prohibited practices | €35M or 7% of worldwide turnover |
| Article 50 and most other duties | €15M or 3% |
| Misleading information to authorities | €7.5M or 1% |
Large firms face whichever is higher; SMEs and startups whichever is lower.
4. Beyond Europe
Companies outside the EU are covered if they place AI on the EU market or their output is used there — Korean, American and other game, media and chatbot firms included. South Korea's AI Basic Act, in force since January 2026, has a parallel rule: services must tell users in advance that they use generative AI and label generated output, with clearer labeling for realistic deepfakes. Both point the same way, but the EU spelled out the "machine-detectable" technical bar first.
The limits are real. Marks can be stripped; in OpenAI's tests, changing a quarter of the words cut watermark detection to 17%. Article 50 is best understood not as a net that catches all AI text but as a rule that makes labeling the default for the honest majority.
5. What remains unconfirmed
- Which marking techniques fully satisfy paragraph 2 depends on the final code of practice and guidance.
- The December 2 legacy deadline is as reported.
- When Korean fines for labeling breaches will actually start was not confirmed.