EU AI Act, Aug. 2: who must label AI content, and who doesn't

News: On July 24, 2026, Google announced it is signing the EU AI Act Code of Practice on Transparency of AI-Generated Content, and says it is working with Apple, Eleven Labs, Kakao, NVIDIA and OpenAI on interoperable watermarking tools built on SynthID. The announcement lands days after the European Commission adopted its final guidelines on Article 50, whose obligations start applying on August 2, 2026 (Google, The Keyword, July 24, 2026).

Direct answer: Article 50 of the EU AI Act starts applying on August 2, 2026, and it creates two distinct obligations that get conflated constantly. The first, Article 50(2), requires providers of generative AI systems to mark their outputs in a machine-readable format, detectable as AI-generated. That is OpenAI's, Google's or Mistral's job, not the job of the business using their tools. The second, Article 50(4), targets deployers: it requires disclosing that a text was artificially generated, but only where it is published for the purpose of informing the public on matters of public interest, and it does not apply where the content underwent human review or editorial control with someone holding editorial responsibility. For most SMEs publishing commercial content reviewed by a human, the labelling duty never triggers. What does change is that editorial review becomes a documentable legal requirement rather than merely good practice.

What Google signed, and what it doesn't say

The announcement was published on the company's official blog by Karen Massin, who leads Google's government affairs and public policy for EU institutions. Google highlights SynthID, its digital watermarking technology, and says it is working with Apple, Eleven Labs, Kakao, NVIDIA and OpenAI toward industry-wide, interoperable adoption.

One point deserves immediate clarification, because it is routinely misreported. A code of practice is not an extra law: it is a voluntary instrument offering a mapped route to demonstrate compliance with the marking obligations. Signing it exempts you from nothing and adds nothing. What binds is the regulation itself, and its application date is August 2. So the announcement is mostly a calendar signal: the large providers are falling into line, which makes technical marking increasingly routine. We saw the same movement when OpenAI joined the C2PA standard and adopted SynthID for its images.

Article 50: two obligations that keep getting mixed up

The reading doing the rounds this week, that "everyone will have to label AI content," is wrong. The regulation draws a sharp line between two roles, and most businesses are touched only by the second, in a narrow set of cases.

In its official FAQ, the European Commission states that three cumulative conditions must be met before the text obligation bites: it must be published, it must inform the public, and it must concern a matter of public interest. It places politics, justice, fundamental rights, public health, environmental protection and consumer safety in that last category. A software comparison page does not tick the third box.

The part nobody says: the editorial exemption describes E-E-A-T exactly

This is where the text gets interesting for anyone producing content. Escaping the disclosure duty is not a checkbox exercise. The Commission defines human review as a deliberate examination of the substance of the content by one or more natural persons possessing relevant knowledge and professional judgement, and editorial control as that of a responsible entity with the authority to approve, alter or reject the substance of the text. It adds that superficial, solely formal or procedural checks, such as spell-checking, do not qualify.

Read that definition again with Google in mind. A named person, competent on the topic, who examines the substance, can refuse publication, and whose responsibility is engaged: that is, word for word, what the quality guidelines call expertise, authoritativeness and trustworthiness. The European regulator and the search engine are asking for the same thing, for different reasons. The compliance artifact and the ranking artifact are the same artifact.

The operational consequence is, for once, pleasant: an organisation that already produces content through a real review chain has almost nothing to change. Mostly, it needs to be able to prove it. The one publishing at volume without genuine review discovers, at the same moment, a legal exposure and the reason generative engines don't cite it, a pattern we documented while analysing the pages AI engines cite from outside Google's top 10.

So are you actually affected?

• B2B blog post, product guide, service page: no. No matter of public interest, and ordinary review closes the question anyway.
• Health, consumer safety, environment or rights content published without substantive review: yes. That is the core target of Article 50(4).
• Generated visual realistically depicting a real person: yes, under the separate deepfake obligation.
• Conversational agent on your site: users must know they are talking to a machine. If you deploy it under your own brand, check whether you become a provider under the regulation.
• Spell-checking or light AI rewriting: no. Assistive editing that does not substantially alter the content is explicitly carved out.

What to do now

1. Map your publications and isolate anything touching health, law, environment or safety. That is the only perimeter where the question genuinely arises.
2. Document the review rather than asserting it. Who reviewed, when, what was changed or rejected. A simple editorial log is enough, but it has to exist before anyone asks for it. It is one of the checks in our GEO audit method.
3. Name an identifiable editorial owner and surface them on the pages concerned. It is the explicit condition of the exemption, and incidentally an authority signal for engines.
4. Audit your conversational agents and the generated visuals already live, which fall under different obligations than text.
5. Don't slap on an "AI-generated" label reflexively. Outside the scope of Article 50(4) it offers no legal protection. Google, for its part, says it judges content on quality regardless of how it was produced: the label protects nothing and improves nothing.

Our take

The AI Act does not ban content made with AI. It makes content made with nobody behind it expensive. That is the same fault line Google has been drawing for three years, arriving this time through the regulatory door, following what California had already set in motion. August 2 will not move your traffic. But it makes enforceable a distinction many publishers preferred to keep blurry.

There is a second-order effect worth naming. Because the EU AI Act pins the exemption to a named person who can reject a text, it pushes editorial accountability back into org charts that had quietly removed it. Teams that scaled publishing by removing the reviewer now have to reinstate the role, and the cheapest way to do that is to publish less and review properly. That is the opposite of the direction most content operations have taken since 2023, and it is the part of the EU AI Act that will actually change behaviour, well beyond the narrow set of pages that legally require a label.

What this article does not cover

This is editorial analysis, not legal advice: for a specific case, have your situation validated by a qualified lawyer. Several areas remain open as of publication. The provisional "AI Omnibus" agreement reached in May 2026 would grant additional time, to December 2, 2026, for machine marking of systems already on the market: it was still provisional when we published, and we do not present it as settled. The consolidated list of signatories to the code of practice is not available. Finally, the designation of national supervisory authorities and their enforcement doctrine are not settled, which makes it impossible to say today how these obligations will be policed in practice in any given member state. We will update this article as those areas firm up.

Frequently asked questions

Alexis Dollé

CEO & Founder

Growth and SEO content strategist, I founded Cicéro to help businesses build lasting organic visibility, on Google and in AI-generated answers alike. Every piece of content we produce is designed to convert, not just to exist.

LinkedIn