The Watermark You Cannot See
Europe just made AI announce itself. Here is what the new transparency law asks of you, and how we rebuilt our own platform to answer it.
By Mogens Rye, Founder of MyAvatar, and Claude AI. August 2026 For most of the short history of generative AI, synthetic content had one quiet advantage: it could pass. A voice that sounded like a person. A presenter who looked like one. A product shot no camera ever took. Half the appeal was that you could not tell. Since the second of August, in Europe, that advantage comes with a rule. AI content now has to say what it is.
That is Article 50 of the EU AI Act, the transparency chapter, and it went live this month. I have spent much of the last fortnight rebuilding parts of my own platform to meet it, so I am not writing this from the sidelines. I am writing it with the tools still on the bench. Here is what the law says, where people are getting it wrong, and what you should do this week if your business publishes AI content.
Article 50 comes down to one idea: when people meet AI, they should know it. It covers four situations. A chatbot has to tell you it is a machine. A system reading your emotions or sorting you by biometric data has to say so. AI-written text published on serious public matters has to be labelled. And if an image, audio clip or video is a deepfake, that has to be disclosed. Deepfake is where most executives relax, and that is the mistake. In ordinary speech, it means a malicious face-swap of a politician or a scam call faking a CEO. The law means something much wider: AI-generated or manipulated content that resembles a real or realistically plausible person, object, place or event, and could appear authentic. Intent to deceive is not required. A photorealistic AI presenter counts. So does a synthetic fashion model or a generated product scene. If it looks real, the law treats it as something the viewer deserves to be told is not. The test is not whether you meant to fool anyone. It is whether someone glancing at the screen might believe it was real.
Two labels, two audiences This is the part that took me a while to get straight, and getting it wrong is what costs money. The law asks for two kinds of marking, aimed at two different audiences. The first is a machine-readable mark: data baked invisibly into the file so other computers can read it. A viewer never sees it. Its job is to let platforms like YouTube, Meta and Google recognize that a file is AI on upload and handle it correctly. This is the provider’s duty, meaning whoever built the system that made the content.
The second is a visible label: something a person can see or hear without any tool. A small “AI-generated” mark on the video, or a spoken line. This is the deployer’s duty, meaning whoever publishes the content. That is you, the moment you post an AI video on your company page. Most people assume the invisible mark covers them. It does not. A hidden machine-readable mark from the provider does not satisfy the deployer’s obligation to a human viewer. One reassures the machines. The other respects the audience. You need both, and they do not stand in for each other. The word “creative” is not a loophole There is one exception, and I want to be blunt about it, because I have watched people reason their way into a false sense of safety.
Deepfakes that are part of an obviously artistic, satirical, or fictional work get lighter treatment. The disclosure can sit in the credits rather than across every frame, so it does not wreck the piece. Fine for a film or a clear parody. The catch is that the exception is narrow, and it does not cover commercial content. Marketing is not art in the eyes of this law. Where a piece mixes commercial and creative character, the commercial side wins, and normal labeling applies. So the “it’s really more of a creative piece” argument a marketing team might reach for does not survive contact with the rule. If you are advertising, you are labeling. The deadlines, and why we ignored the easy one.
The timing has a nuance worth knowing.
The visible deepfake disclosure, the deployer’s job, applies now, from the second of August, with no grace period. The invisible machine-readable mark gets a short extension: for systems already on the market before that date, it is required from the second of December 2026. I read that extension as breathing room, not permission to wait. The visible duty is already live, and the penalties behind the whole thing are real: up to 15 million euros or 3 percent of worldwide annual turnover, whichever is higher, with the lower figure applying to smaller companies, enforced by national authorities across all 27 member states. Looking at those numbers, I decided to build everything now rather than lean on December. So we did.
What we built Here is the workshop-floor part. MyAvatar produces video: you write a script, the platform returns a finished piece with an AI presenter, ready to publish. That puts us in an unusual double position under this law. We are the provider, because our engines generate the content, so the invisible mark is on us. And when we publish our own videos, we are a deployer, same as any customer. So we built both layers into the single point every finished video passes through on its way out. For the invisible layer, we took two open standards rather than rolling our own. One is C2PA, a tamper-evident seal of provenance: a signed note inside the file recording what made it and when. The other is a neural watermark, a pattern woven into the pixels themselves and built to survive re-compression and re-uploading. One is a signed label on the file; the other is written into the image. Together they mean a MyAvatar video declares its origin to any machine that checks, and keeps declaring it after the video has been shared, clipped and re-posted.
For the visible layer, we burn the “AI-generated” mark into the video itself, in the viewer’s language, kept on screen for the whole clip rather than flashed once at the start. On social media, people scroll in halfway. A label only the early viewers see is not really a label. We treated this as part of how a video is born, not a sticker applied at the end. One honest caveat, because this series does not do hype: no marking system is unbreakable. A determined bad actor with the right tools can degrade a watermark, and detecting manipulated content is an arms race the defenders do not always win. What these layers do is make misuse harder and honest content verifiable. That is the real goal. Anyone promising you a hundred percent is selling you something. Where my thinking flipped
Somewhere in the middle of the rebuild, my view of the whole thing changed. I had been treating Article 50 as a tax: a cost of operating in Europe, a box to tick. By the end I saw it as the opposite. For a customer who cannot afford a single misstep, like a children’s brand, a bank, a hospital or a public body, provable transparency is not a burden. It is the thing they have been waiting for. A toy company can now show a wary parent exactly what in an advert is AI and what is real, with proof rather than a promise. That is not a compliance headache. It is a reason to pick you. And there is a small loop in all this that I enjoy. This letter will become a video on my own platform. A piece about AI transparency, made by an AI system, carrying the very marks it describes. It will tell you, in plain sight, that it is AI. Which is the whole point.
What to do about it One concrete move for this week, small enough to actually finish. Take an inventory. List every place your organization publishes AI-generated or AI-edited content: the videos, the generated ad images, the synthetic voiceovers, the AI-written posts. Against each one, ask two questions. Does a human viewer get told it is AI? Could a platform detect it automatically? Wherever the answer is no, you have found a gap. You do not have to close all of them this week. You have to know where they are, because most companies do not, and the ones who find out first are the ones who avoid the nasty surprise. The rule is not coming. It is here. The only question left is whether your content is honest by design or honest by accident.
About the author Mogens Rye is the founder of MyAvatar, an AI-native video production platform based in Copenhagen. He is a former stockbroker turned entrepreneur, and has built MyAvatar’s multi-engine platform as a partnership between a single human founder and a team of specialized AI agents. MyAvatar’s platform is available at myavatar.dk.