On 2 December 2026 one of the most practical rules in the EU AI Act reaches the companies that were given extra time. Article 50(2) says that AI systems which generate synthetic audio, images, video or text must mark that output in a machine-readable format, so that it can be detected as artificially generated.
Article 50 as a whole has applied since 2 August 2026. The Digital Omnibus (Regulation (EU) 2026/1744) gave generative AI that was already on the EU market before 2 August 2026 until 2 December 2026 to comply with the marking duty. The same day, two new prohibited practices under Article 5 start to apply.
The question we hear most is simple: is this date mine?
It binds providers, not everyone who uses AI
The marking duty in Article 50(2) sits with the provider: the company that develops a generative AI system, or has it developed, and offers it under its own name. In practice that means you if you:
- sell or license a SaaS feature that writes text, creates images or produces audio or video
- ship an app or a chatbot that generates content for your customers
- built a content generator on top of a model and offer it to others under your brand
If your company only uses AI tools (a writing assistant, an image generator, a chatbot platform you subscribe to), the marking duty is your vendor’s, not yours. That does not leave you with nothing to do. As a deployer you have had Article 50 duties of your own since 2 August 2026:
- Deepfakes: if you publish image, audio or video content that is AI-generated or manipulated and looks real, you disclose it.
- Public-interest text: AI-generated text published to inform the public on matters of public interest is disclosed as such, unless it went through human review and someone holds editorial responsibility for it.
- Emotion recognition and biometric categorisation: if you use these systems, you inform the people exposed to them.
Not sure which side you are on? We wrote a separate note on provider or deployer, because the answer decides most of your obligations.
What “machine-readable marking” means in practice
The Act does not prescribe one technology. It asks for technical solutions that are effective, interoperable, robust and reliable as far as technically feasible, taking into account the type of content and the state of the art. The recitals name the usual families:
| Technique | What it does | Good for |
|---|---|---|
| Watermarks | A signal embedded in the content itself | Images, audio, video |
| Metadata | Provenance information attached to the file, such as signed content credentials | Files you export and share |
| Cryptographic methods | Signatures that prove where the content came from | Proving origin, detecting tampering |
| Logging and fingerprints | A record of what was generated, to match content later | Text, where watermarks are weaker |
Most providers end up with two layers: one that travels inside the content and one that travels with the file. Text is the hardest case, because short text loses most signals when it is copied, so logging and the model vendor’s own mechanisms matter more there.
There is one exception worth knowing: the duty does not apply where the AI performs an assistive function for standard editing, or does not substantially alter the input data. A grammar fixer is not a content generator.
Six weeks of work, in order
If the date is yours, this is the order we follow with clients:
- List every output channel. Text in the app, images in exports, audio in downloads, video in the API. Marking has to cover each one, not just the main screen.
- Check what your model vendors already do. Some foundation-model providers mark output at the source. Know what you inherit, and do not strip it out in your pipeline.
- Add your own layer where the vendor’s is missing. Usually metadata or signed content credentials on exported files.
- Do not let your terms or your code remove the marks. Resizing, compression and format conversion can strip metadata without anyone noticing. Test it.
- Write it down. A short technical note: what you mark, how, where it can fail, and how you test it. If anyone asks, that note is your answer.
- Handle Article 50(1) at the same time. If people talk to your system, they need to know it is an AI. That duty has applied since August.
What we do not know yet
The Commission is expected to publish guidance and a code of practice on transparency for AI-generated content. Until then, the safest reading is the text of the Regulation itself, applied with judgement. If you have read a different date somewhere, check it against Regulation (EU) 2026/1744: several public trackers still showed the pre-Omnibus calendar when we last checked.
Sources. Regulation (EU) 2024/1689, Article 50 and recitals 133 to 136, as amended by Regulation (EU) 2026/1744 (Digital Omnibus on AI). Dates verified on 5 October 2026.
A practical readiness and implementation engagement. It does not replace your legal counsel, a formal conformity assessment or certification. We work alongside your counsel, never in its place.