The code of practice on transparency of AI content
The code is voluntary and is not law. Signing it lets you demonstrate compliance with Article 50(2) and (4) AI Act. Without signing, the same thing has to be demonstrated another way.
Article 50 AI Act says that content has to be marked. What that looks like technically, the regulation does not say. The code of practice on transparency of AI-generated content is meant to close that gap. Article 50(7) AI Act foresees it as a Union-level practice guide. After drafts in December 2025 and March 2026, the published version has been available since 10 June 2026.
Two sections, two audiences
It addresses providers of generative AI systems and covers machine-readable marking and the detectability of generated or altered content, meaning Article 50(2) and (5) AI Act. It recommends a multilayered approach combining embedded metadata, an embedded watermark and supplementary fingerprinting. The reason for the layering is simply that none of the techniques suffices on its own. Watermarks can be removed, metadata is regularly lost when uploading to platforms, and for text there is still no market-ready solution.
It addresses deployers and covers the marking of deepfakes and of AI-generated or altered text, meaning Article 50(4) and (5) AI Act. Here the question is less about technology than about the timing and form of disclosure. It has to happen at the first encounter at the latest, be conspicuous and comprehensible, and stand out from its surroundings. A note in the credits or several menu levels down does not meet that.
Sign, or prove it yourself
The real incentive sits in Article 50(7) AI Act. Anyone following a code that the AI Office has assessed as adequate can use it to demonstrate compliance with the duties in Article 50(2) and (4). The duty itself does not change. The evidential position does.
Having signed
- Compliance is demonstrated through the code
- Supervision focuses mainly on whether you actually work to it
- The technical requirements are given and do not have to be derived
- Changes in the state of the art arrive through the code
Without signing
- The duty remains exactly the same
- You produce the evidence yourself, usually as a documented gap analysis
- Every departure from the code needs reasons, not just a mention
- Expect more frequent requests for information and access
What this means internally
Marking is not a task for the legal department alone. It reaches into editorial work, marketing and customer service, and it only works if it is built into the existing approval processes rather than sitting beside them. Anyone reusing output also has to make sure an existing marking survives the process. That is the duty deployers overlook most often, because it is broken incidentally when an image is cropped or a video converted.
The dates are not uniform. Article 50(1), (3) and (4) AI Act apply from 2 August 2026. For machine-readable marking under Article 50(2) the Digital Omnibus created a transitional rule that benefits only systems already offered before that day. For those the period runs until 2 December 2026.
How we help
We compare your marking practice against the code, record the departures with reasons, and draft the internal rule for it. Where signing is an option, we say what it actually means in day-to-day operation before it is declared.
Generated content in circulation, marking unsettled?
We check what the code requires for your channels and how you produce the evidence.
Get in touchFrequently asked questions
What is the Code of Practice on Transparency?
A Union-level practice guide foreseen by Article 50(7) AI Act. The published version is dated 10 June 2026. Its first section addresses providers of generative systems and covers machine-readable marking. The second addresses deployers and covers deepfakes and AI-generated text.
Do we have to sign it?
No. It is voluntary and does not replace the regulation. Signing does more than a recommendation would, though. Under Article 50(7) AI Act, compliance with the duties in Article 50(2) and (4) can be demonstrated by following a code that the AI Office has assessed as adequate. Supervision then focuses mainly on whether the signatory actually keeps to it.
What if we do not sign?
The duty stays the same, only the proof becomes more work. The usual route is a documented gap analysis setting your own measures against the requirements of an adequate code and giving reasons where you depart from it. Anyone who neither signs nor produces such evidence should expect more frequent requests for information and access from the supervisory authority.
Which technique does the code require for marking?
It does not settle on one but recommends a multilayered approach. Metadata under the C2PA standard, an embedded watermark and supplementary fingerprinting are meant to work together, because none of them currently suffices on its own. Whether this approach closes the structural gaps that follow from the probabilistic nature of model output is disputed in the literature.
From when do the duties the code relates to apply?
The duties in Article 50(1), (3) and (4) AI Act apply unchanged from 2 August 2026. Machine-readable marking under Article 50(2) has a transitional rule. Systems placed on the market after 2 August 2026 have to meet it from that day, while systems already offered before then have until 2 December 2026. Output generated and published before 2 August 2026 does not have to be marked retrospectively.
Are there consequences beyond fines?
Yes, and they are frequently underestimated. Breaches of Article 50 AI Act can be fined up to 15 million euros or three per cent of worldwide annual turnover. Alongside that stands unfair competition law. The prevailing view treats the transparency duties as market conduct rules, which would make a breach actionable under section 3a UWG. Independently of that, misleading omission under sections 5a and 5b UWG comes into play. Both open the door to warning letters from competitors, and those follow no regulator's timetable.