From August 2, 2026, all AI-generated content distributed in Europe must be marked as such, in a machine-readable manner. The same day, the same rule becomes operative in California.
From August 2, 2026, all AI-generated content distributed in Europe must be marked as such, in a machine-readable manner. The same day, the same rule becomes operative in California. Four major markets now require this marking. Only one has not made the technical method restrictive: ours.
What actually changes on August 2
Do you operate a chatbot, image generator or voice assistant accessible in the EU? Article 50 of the AI Act imposes, depending on your role in the chain, two families of obligations: telling the user that he is talking to an AI if you provide the system that interacts with him, and embedding in each generated output a machine-detectable marking if you provide the system that generates it. The Commission published its guidelines on July 20 to frame all of this.
The addition in the event of a breach is written in article 99: up to 15 million euros or 3% of global turnover, the lower of the two for SMEs. Firm obligation, firm date, firm fine. There remains the question that every engineer asks first: how do I implement it?
The problem: an obligation without enforceable spec
The official response is contained in a code of good practice published on June 10, 2026. It describes how to mark, detect, label. Except that this code is voluntary. No one is obliged to sign it.
Said in tech language: Europe has shipped the obligation in production without freezing the spec. The code exists in final version, everyone will refer to it, no one is required to do so.
It will be objected to me that the spec already exists: C2PA and Content Credentials, whose steering committee brings together Adobe, Google, Microsoft, OpenAI and Sony. This is precisely the problem. A standard governed by a private industrial consortium is the opposite of an enforceable public benchmark: useful as a technical building block, it cannot serve as a text against which a European authority judges conformity. Unless we admit that the examination room belongs to the examinees.
And the resulting mechanism is explicit in the Commission itself: the signatories of the code will be monitored on their adherence to the code. The others will have to prove, on a case-by-case basis, before each national supervisory authority, that their in-house solution does the job. An optional text becomes the de facto reference of a legal obligation. It’s comfortable for the signatories. It’s soft law for everyone.
Elsewhere, the spec is in the law
In China, marking measures, in force since September 1, 2025, impose the combo of visible label plus embedded metadata, and expressly refer to mandatory national standards. We can criticize everything about Chinese control of information, and we must. But the architecture is coherent: the obligation and the method are both binding. South Korea has made a similar choice, framework law and implementing decree in force since January 22, 2026 with labeling of generative AI content. Useful clarification: an opposable reference system is a necessary condition of proof, not a guarantee of application, and the Korean press is already documenting the gap between the two.
In California, the text of the law AB 853 also becomes effective on August 2, 2026. Large generative AI providers must onboard provenance marks and publish a free and open detection tool, or face $5,000 per violation. Remember the design: the law does not only impose the watermark, it imposes the detector which allows anyone to check it. The obligation embeds its own proof, like a test that ships with the code.
The most surprising: the fix already exists in the regulations
Article 50, paragraph 7 organizes two routes: the Commission can approve a code of good practice by implementing act, and if it judges the code inadequate, impose by implementing act common binding rules for marking and detection. The Commission and the European AI Committee have deemed the code suitable as a voluntary tool. The binding route therefore exists in the regulation, it remains legally subordinate to a finding of inadequacy which the Commission has just set aside. In other words: the mechanism of opposability is written, and the institutional choice which has just been made keeps it at bay for the long term.
Let’s summarize the landscape of August 2: Europe activates the heaviest fine of the four markets with the weakest technical benchmark of the four. An obligation is proven against a spec enforceable against all, not against a document to which everyone chooses to adhere. The real question for a European product team is not whether it will be able to brand its content. It is to know against which text, and before whom, she will prove that she is doing it correctly.
What to do now
Business side. Align yourself with the code of good practice even without signing it: today it is the only text against which a supervisory authority will assess your measures, and the Commission has written in black and white that alternative solutions will be judged on a case-by-case basis. Document each implementation choice (tagging format, metadata, robustness) as a proof file, not a technical note. What is not documented does not exist before an authority.
European side. The first review of the system should ask the question that the regulation already asks: if the voluntary code is sufficient, why did Article 50(7) provide for the route of a binding implementing act? We will respond that the technique is not mature, that the watermarks are being removed and that to freeze a spec would be to impose the impossible. The argument proves the opposite of what it thinks it proves: immaturity pleads for an evolving reference, revisable by act of execution, not for an optional reference. California shows that opposable and democratic are compatible. Europe has the mechanism. All he needs is the decision to use it.
