The European Commission has finalized its transparency rules for AI-generated content, and the obligations start to bite on August 2, 2026. The rules sit under Article 50 of the EU AI Act and come with a Code of Practice that spells out how providers and deployers should disclose synthetic content. PPC Land reported the details. Penalties for getting it wrong reach up to €15 million or 3% of worldwide turnover.
What the rules actually require
The obligations split by role. Providers, the companies that build and supply generative AI systems, have to machine-mark the content their tools produce so it can be detected as AI-generated. For text, the marking requirement kicks in once a passage runs longer than 200 tokens, roughly a few paragraphs. That marking has to be readable by other software, which is what makes downstream detection possible.
Deployers, the businesses and people who use these systems to publish, carry a different duty. They have to add a visible label when they put out deep fakes or AI-generated text on matters of public interest. So an image or video that has been synthetically produced or altered needs a clear notice, and so does AI-written text that speaks to public-interest topics.
There is one exemption worth reading twice. AI-generated text that has gone through human editorial review and fact-checking, with a person or organization taking responsibility for it, does not need the visible label. That carve-out matters for anyone running a real editorial process rather than publishing raw model output.
The timeline and the money
Enforcement starts August 2, 2026. Systems already on the market get more runway: they have until December 2 to conform. The fines are the part that will focus attention. Up to €15 million or 3% of worldwide annual turnover is a number that reaches the board.
Watermarking, in plain terms
Two different things are happening here, and they are easy to confuse. Machine marking is invisible to a reader but detectable by software, and it applies to what providers generate. Visible labeling is the human-readable notice a deployer adds on deep fakes and public-interest text. A marketing team can be on the hook for the second even when a vendor handles the first, so it helps to know which side of the line each task falls on.
For text specifically, the 200-token threshold is the trigger. Short snippets sit below it. Anything approaching article length is above it, which covers most blog posts, landing-page copy, and long social captions produced with AI assistance.
What this means for Thai marketers
The AI Act applies based on who the content reaches, not where the agency sits. A team in Bangkok producing AI-assisted content for a client that sells into the EU, or publishing on a global platform with European users, falls within scope. The practical move is to build labeling and watermarking into the workflow now rather than retrofitting it after August.
A few things are worth sorting out early. Know which of your content is machine-marked at the source by the AI vendor, and which needs a visible label that your team applies. Decide where AI text touches public-interest topics, since that is where the visible-label duty lands. And formalize the editorial review step, because a documented human check with fact-checking is what unlocks the exemption for AI-written text. Strong content marketing processes already include an editing pass, so this is often a matter of writing down what good teams do anyway.
If AI is part of how you produce or optimize pages, the disclosure question also touches AI SEO work, where model-generated copy and structured content are common. Treating transparency as a standard field in the brief, rather than a compliance afterthought, keeps the December deadline for existing content from turning into a scramble.
FAQ
When does enforcement start?
August 2, 2026. Content and systems already in market have until December 2, 2026 to conform.
Does every AI text need a visible label?
No. The visible label applies to deep fakes and AI text on matters of public interest. AI text that has had human editorial review and fact-checking is exempt.
How large are the penalties?
Up to €15 million or 3% of worldwide turnover.
If your content reaches European audiences and you want a workflow that handles disclosure without slowing production, Relevant Audience can help you set it up. Talk to us about building it into your content and AI SEO process.
