On 10 June 2026 the European Commission published the final Code of Practice on marking and labelling AI-generated content. The code itself is voluntary; the obligations it helps you meet are not. Those rules, set by Article 50 of the EU AI Act, become applicable on 2 August 2026, and they split cleanly into two jobs: the companies that build AI tools must mark their output, and the people who publish AI content must label it. This guide explains which side you are on and what to do before the deadline. It builds on our 2026 AI content compliance guide and our explainer on C2PA content credentials.
What is the EU's AI content labelling Code of Practice?
It is a playbook from the European Commission, published 10 June 2026, for meeting the transparency duties in Article 50 of the AI Act. Signing it is optional, but the underlying law applies whether you sign or not. The code spells out, in practical terms, how to mark and disclose synthetic media so that a viewer and a machine can both tell the content was made or altered by AI.
Who must do what, and by when?
Two roles, two obligations. Read the row that describes you.
| Role | What you must do | When |
|---|---|---|
| AI provider (you build a generative tool) | Embed machine-readable marks: digitally-signed metadata + an imperceptible watermark; offer a free public detector (spec, software, or API) | 2 Aug 2026 · marking for tools already live before that date deferred to 2 Dec 2026 |
| Deployer (you publish AI content) | Visibly label deepfakes and AI-generated text on matters of public interest, using the EU AI Office's standard icons | 2 Aug 2026 |
Does this apply to creators or only to AI companies?
Both, in different ways. If you build a generative model or tool, you carry the provider duties: watermark the output and make a detector available. If you use AI to publish, which describes most creators, you carry the deployer duty: label the synthetic content you put out. A faceless channel running AI voiceover and a marketing team shipping AI articles both fall on the deployer side.
What exactly does a publisher have to label?
Two things: deepfakes and AI-generated text on public-interest topics. A deepfake here means realistic synthetic audio or video of real people or real events, the kind a viewer could mistake for genuine footage. AI text falls in scope when it covers matters of public interest, such as news or politics. The disclosure has to be visible, and the Code points to a set of standard icons from the EU AI Office so the label looks consistent across the web. Routine AI-assisted editing that does not mislead is not the target.
What about the watermarks and the hidden metadata?
That part lands on the tool makers, but it affects you. Providers embed two kinds of signal into every output: digitally-signed metadata and an imperceptible watermark, plus a free way for anyone to check a file against their system. The practical takeaway for a creator: assume the AI media you generate already carries a hidden, hard-to-remove mark, and that platforms can read it on upload. This is the same provenance logic behind SynthID and C2PA content credentials, now backed by EU law.
When do the rules actually bite: 2 August or 2 December?
Both dates matter. The core Article 50 transparency duties, including disclosing AI interaction and labelling synthetic content, apply from 2 August 2026. The narrower requirement to embed machine-readable marking gets a grace period for tools already on the market before that date, pushed to 2 December 2026. Penalties for Article 50 breaches reach up to €15 million or 3% of global turnover, so the dates are not soft.
Is the Code mandatory?
The Code is voluntary; the obligations behind it are law. Signing gives a company a clear, agreed way to show it complies, which works like a safe harbour. Skipping the Code does not skip Article 50: a non-signatory still has to meet the same transparency duties by some other route, just without the ready-made template.
What should a creator do now?
- If any of your audience sits in the EU, plan to label realistic AI video, audio, and public-interest AI text with clear, visible disclosure.
- Keep provenance signals on, such as Content Credentials, wherever your tools support them.
- Do not strip watermarks or metadata from AI output; that defeats the marking and invites trouble.
- Mark the calendar: 2 August 2026 for transparency duties, 2 December 2026 for machine-readable marking on older tools.
The bottom line
A voluntary code sits on top of mandatory law, and the clock runs to 2 August 2026. Tool makers must watermark and offer a detector; publishers must visibly label deepfakes and public-interest AI text. If you serve an EU audience, start disclosing realistic AI content now rather than scrambling at the deadline. For the wider picture, start with our 2026 AI content compliance guide.
