Last Sunday, one of the most talked-about obligations of the European Artificial Intelligence Act came into effect. As of 2 August 2026, anyone publishing an image, audio or video in the European Union that has been generated or manipulated with AI and could pass for authentic must state this clearly and perceptibly. Furthermore, whoever manufactures the tool must technically label what it produces, so that platforms and fact-checkers can recognise it automatically.
This is good news. It is worth saying, as we have been asking for precisely this for years.
It is also worth mentioning the other thing: if a video of you appears tomorrow saying something you never said, that video will not carry a label. And it won't carry one precisely for the same reason it exists.
What exactly does Article 50 mandate
It is worth understanding the standard before discussing it, as quite a lot of confusion circulates.
Article 50 of Regulation (EU) 2024/1689 does not require a notice to be stamped on everything made with artificial intelligence. It is triggered in four specific situations: when a system interacts directly with people — hence chatbots must identify themselves as such —, when emotion recognition or biometric categorisation is used, when synthetic content is generated, and when that content constitutes a deepfake, defined as material that resembles real people, objects, places or events and which could appear authentic.
And it works on two levels that constantly get confused.
The first falls on the provider: to mark the file in a machine-readable format. It is a technical mark, invisible to you. Systems that were already on the market before 2 August have until 2 December 2026 to comply with this specific part.
The second falls on whoever deploys the system and publishes the result: to inform in a clear, distinguishable and comprehensible way that it is artificial. A hidden watermark inside the file is not enough. The person exposed to the content has to find out at the moment of seeing it.
Fines for failing to comply with these transparency obligations can be as high as 15 million euros or 3 % of global turnover. The Commission published its final guidelines on 20 July, and the so-called Digital Omnibus, agreed in May, delayed the application of the high-risk obligations but did not affect Article 50. This deadline has now passed.
The recipient problem
Here is the nub of the matter, and it is the same one that is repeated with almost all internet regulation.
The regulation speaks to those who are going to comply with it. It speaks to the advertising agency that generates a campaign with AI, the media outlet that illustrates a news story, the brand that deploys a customer service chatbot, and the studio that produces an audiovisual piece. All of them have a name, a registered address, legal counsel and something to lose. They are going to label.
Whoever creates a fake sexual video with your face, clones your voice to ask your chief financial officer for a bank transfer, or fabricates a statement you never made to ruin you in a negotiation or a custody battle, has no tax domicile and no intention of complying. They post from an account created that morning, on a hosted platform where a legal request takes months, and then disappear.
To put it awkwardly: Article 50 raises the floor of the legitimate market, which was necessary. It doesn't disarm someone who wants to do you harm.
What does change for you
And yet, something changes. Only not in the way it seems.
When labelling becomes the norm, not labelling ceases to be neutral. Until now, synthetic content without a warning was simply synthetic content without a warning. As of this week, it is a breach. And a breach serves a very specific purpose in a claim: it reinforces the intentional element, the one that is always the hardest to prove when someone claims it was a joke, a parody or an experiment.
Change the conversation with the platforms too. A notification that says “this is fake” competes with the word of the person who uploaded it. A notification that says “this is unidentified synthetic content, in a framework where identification has been mandatory since 2 August 2026” is a different argument, and moderation teams treat it differently.
And the criminal landscape in Spain is changing, although not entirely yet. The Organic Bill for the protection of minors in digital environments, currently going through parliamentary processing, plans to reform the Criminal Code to explicitly criminalise sexually explicit deepfakes created with a person's face or body without their consent. In parallel, in May, following a proposal by Spain, the European Union agreed to directly ban AI systems intended to generate that type of material.
If yours is already on the road
None of the above works when the video is already in a WhatsApp group of two hundred people. There, what you have to do is something else, and you have to do it in order.
First, document before it moves: dated screenshots, URLs, account identifiers, and, if possible, a download of the original file, because the metadata and the provider's eventual technical watermark may be precisely the proof you need later. Publishing a denial before having preserved the evidence is a mistake that costs dearly.
Then, act on two fronts at the same time. On the one hand, removal, supported by the Digital Services Act, which obliges platforms to have notice and action mechanisms, and on the right to erasure against any copies that may have been replicated. On the other hand, the deindexing, because the video might disappear from the original platform and still appear when someone types your name into a search engine. They are two different jobs and neither replaces the other.
And in parallel, assess the civil route for intrusion into honour and self-image, or the criminal complaint depending on the content. With a warning that we have already made here and which is relevant once more: the clock start running earlier than people think.
A note for businesses
This article will be read mainly by people worried about being the target of a deepfake. But there is a second reading that is worth not overlooking.
If your company uses generative AI to produce content—product images, voiceovers, videos, a web assistant—you have had concrete obligations since Sunday. And the risk is not just the fine. It is that a competitor, a disgruntled customer or a former employee might publicly point out that you are publishing unidentified synthetic material. That is a textbook reputational crisis, with the particularity that in this case they would be in the right.
Doing things right and being able to prove it is no longer just a compliance formality; it has become an asset of trust.
If you are in either of the two situations, we will review your case free of charge, with no obligation and in complete confidence. We are just a click away.




