Do You Have to Label AI-Generated Content in the EU? Articl…

Since 2 August 2026, the EU AI Act requires AI-generated content to be marked. Most marketing teams have read that as "we now have to put an AI disclaimer on everything." That is not what Article 50 says.

The article splits the work between two parties, and the split matters:

Your AI vendor has to embed an invisible, machine-readable mark in what its model produces. This happens automatically. You cannot see it and you do not have to do anything.
You, as the business publishing the output, have to add a visible disclosure only in specific situations: chatbots, deepfakes, and AI-written text published to inform the public on matters of public interest.

For most marketing output, a product page, a newsletter, an ad, a social post, no visible label is required. But the invisible mark is already in your content, whether you label it or not. That is the part worth understanding.

What Changed on 2 August 2026

The transparency obligations in Article 50 of the AI Act became applicable and enforceable by national authorities on 2 August 2026.

This date survived the Digital Omnibus on AI (Regulation (EU) 2026/1744, in force 27 July 2026), which pushed the high-risk AI system deadlines out to 2027 and 2028. Article 50 was left untouched. If you were waiting for a delay that applied to transparency, it did not arrive.

The Omnibus did grant one narrow extension: generative AI systems already placed on the market before 2 August 2026 have until 2 December 2026 to meet the machine-readable marking duty in Article 50(2). That is a supplier deadline, not yours. In practice it means some of the AI features in your stack are marking output today and some will start in December.

What Your AI Vendor Has to Do

Article 50(2) puts the burden on providers of systems that generate synthetic audio, image, video, or text. Their outputs must be "marked in a machine-readable format and detectable as artificially generated or manipulated."

Two techniques do the work in practice:

Watermarking. A statistical signal woven into the text or media itself, invisible to a reader, designed to survive copying and light editing. Google's SynthID is the best-known implementation and has been applied to over 100 billion images.
Signed provenance metadata. Cryptographically signed data attached to a file recording how it was made, following the C2PA open standard developed by Adobe, Microsoft, the BBC and others.

Anthropic's implementation is a useful reference because the company documented it publicly. Claude embeds an imperceptible watermark directly in generated text, and attaches C2PA metadata to generated SVG, PNG and JPG files. It covers models launched from 2 August 2026 onward, applies across the API, the web app, Claude Code and cloud partners including AWS, Google Cloud and Microsoft, and it applies worldwide rather than only to EU users.

Two details in that documentation deserve a marketer's attention.

First, marking is not limited to content the model wrote. Anthropic states that proofreading, translation, summarising and file conversion may all produce a mark, even when the ideas and the words came from a human. If your editor runs a human-written article through an LLM for a grammar pass, the output may carry a mark.

Second, the vendor's own caveat runs in both directions. A detected mark indicates the model may have processed the content, not that it authored it. And the absence of a mark does not mean content was not AI-generated, particularly if it was heavily rewritten, very short, or produced by an older model.

Who has committed to this

The Commission published a voluntary Code of Practice on Transparency of AI-generated Content on 10 June 2026. Around 190 organisations had signed by the end of July.

Signatories to the provider section include Aleph Alpha, Anthropic, Black Forest Labs, Cohere, Google, Meta, Microsoft, Mistral and OpenAI. The deployer section has been signed by companies including Getty Images, Lenovo, Lufthansa, Iberdrola, Bulgari and Fastweb.

Signing is not mandatory. Providers who do not sign have to demonstrate compliance through other adequate means, which is a harder conversation with a regulator than pointing at a code you adhere to.

What You Have to Do

Three obligations can land on you. Which ones apply depends on what you publish, not on which model you used.

Tell people when they are talking to a bot

Article 50(1) requires that people interacting directly with an AI system are informed of that fact, unless it is obvious from the context. A support widget labelled "AI assistant" clears this easily. A chat bubble presenting itself with a human first name and a headshot does not.

If you run AI chat on your site, check the disclosure copy in your live chat tool rather than assuming the vendor handles it. The provider has to make the capability available. Configuring it is yours.

Disclose deepfakes

Article 50(4) requires deployers to disclose AI-generated or manipulated image, audio and video content that resembles real people, objects, places or events. Clearly fantastical or physically impossible content falls outside the definition, and artistic, creative, satirical or fictional work only needs disclosure in a way that does not spoil the work.

For marketing, the realistic triggers are synthetic spokespeople, AI voice-overs of a real person, and product or location imagery composited to look like a real photograph.

Disclose AI-written text on matters of public interest

This is the obligation most often misread. Article 50(4) covers AI-generated text "published with the purpose of informing the public on matters of public interest."

The trigger is your purpose in publishing, not the topic alone. Commercial and promotional content does not normally qualify. A campaign page, a product announcement, a nurture email and a case study are not attempts to inform public discourse.

There is also an exemption that covers most serious content operations: text that has undergone substantive human review, with a named party holding editorial responsibility, does not need labelling. The Commission's guidance is explicit that this cannot be superficial or a cursory sign-off. Real editorial judgement, not a rubber stamp.

Where each scenario lands

What you publish	Visible label required?
AI-drafted product or landing page copy	No
AI-drafted newsletter or nurture sequence	No
AI-generated illustration, obviously stylised	No
AI chat widget on your site	Yes, disclose it is AI (Art. 50(1))
Synthetic voice or video of a real person	Yes, deepfake disclosure
Photorealistic AI image of a real place or event	Yes, deepfake disclosure
AI-written explainer on policy, health or elections, no human editor	Yes
The same explainer, substantively edited under named editorial responsibility	No

Where disclosure is required, Article 50(5) sets the timing and form: clear, distinguishable, accessible, and provided no later than the first interaction or exposure. A note in your terms of service does not satisfy this.

Four Practical Consequences for Marketing Teams

1. Your AI-assisted content is already marked. Marking is a property of the output, not a setting you enable. Anything drafted, translated or summarised by a recent frontier model may carry a watermark that persists through copy, paste and editing. This is not a legal problem, most marketing copy needs no label, but assume the provenance signal exists.

2. Your image pipeline probably destroys C2PA metadata. Resizing, re-encoding and CDN optimisation strip file metadata, and most social platforms re-encode on upload. That does not put you in breach, since machine-readable marking under 50(2) is the provider's duty and not yours. But it does mean your provenance trail dies at your CMS. If provenance matters to you, keep the original files.

3. Detection tools are evidence, not proof. Watermarks can be removed by aggressive paraphrasing, and a short passage may carry no detectable signal at all. Do not build an agency contract clause, a contributor policy or a plagiarism process on the assumption that a detector settles the question. Vendors say so themselves.

4. Add marking to your vendor questions. When you evaluate an AI marketing tool or an AI search and GEO tool, ask three things: does it mark generated output, has it signed the Code of Practice, and does it preserve provenance metadata on content passing through it. Vendors that cannot answer are relying on their upstream model provider without having checked.

What Non-Compliance Costs

Breaching Article 50 sits in the middle penalty tier of the AI Act. Under Article 99, fines reach 15,000,000 EUR or 3% of total worldwide annual turnover, whichever is higher. For SMEs and startups, the calculation flips to whichever is lower.

Enforcement is by national competent authorities, which is the same structure that produced very uneven GDPR enforcement across member states. Uneven does not mean absent.

This Is Not a European Versus Non-European Question

It is worth being clear, because the framing gets muddled: Article 50 applies based on where the output is used, not where the vendor is incorporated. A US model provider serving EU users is in scope. Anthropic's marking rolls out worldwide, not to EU traffic only.

So unlike Schrems II or CLOUD Act exposure, this is not an area where a European vendor gives you a structurally different legal position. What it does change is practical: a smaller European vendor is more likely to give you a direct answer about how its AI features mark output and whether provenance survives its pipeline, and it is answering under the same regulator you are.

The AI Act also does not replace anything you already owe. If your AI feature processes personal data, you still need a Data Processing Agreement and a lawful basis, and if it profiles visitors you still need valid consent.

Browse verified European tools in the AI marketing, marketing automation and web analytics categories, or explore the full directory.

FAQ

Do I need to label AI-generated blog posts?

Usually no. The Article 50(4) labelling duty covers text published to inform the public on matters of public interest, and it does not apply where the text had substantive human review under clear editorial responsibility. A commercial blog post that an editor genuinely worked on falls outside it on both counts. A high-volume, unreviewed content operation publishing on policy, health, finance or civic topics is a different matter.

Is the invisible watermark the same thing as a label?

No, and conflating them is the most common mistake. The watermark is machine-readable marking under Article 50(2), embedded by the model provider so that automated tools can detect AI involvement. A label is a human-visible disclosure under Article 50(4), which you add. One is invisible and automatic, the other is visible and your responsibility.

Does AI watermarking affect my SEO?

There is no announced search-engine treatment of watermarked content. Google has not said it will demote content because a mark is detected, and its guidance continues to judge content on usefulness rather than how it was produced. Since detection is neither conclusive nor complete in either direction, a ranking system built on it would be unreliable. Treat the marking rules as a compliance matter, not a ranking one.

Does this apply to content I published before August 2026?

No. The Commission has confirmed that content created before the obligations applied does not need retroactive labelling, though it encourages it. Separately, generative systems that were already on the market before 2 August 2026 have until 2 December 2026 to implement machine-readable marking.

Who is liable if my AI vendor does not mark its output?

The marking duty under Article 50(2) is the provider's, so a vendor that fails to mark is the party in breach. Your exposure is separate: it is the visible disclosure duty under Article 50(1) and 50(4) for what you publish and how you deploy chat. That said, a vendor ignoring Article 50 is a signal about the rest of its compliance posture, which is a reason to ask before you buy.