A company that puts an AI avatar into an advert for European audiences has a new legal obligation from 2 August 2026. It must tell viewers that the person speaking in the video is not a human. Article 50 of the European AI Act became enforceable on that day, and the obligation falls directly on whoever publishes AI content.
Until now, the debate about AI regulation in Europe has revolved around models and their providers. Article 50 shifts the focus to distribution. According to the European Commission's official overview, the transparency rules apply from 2 August 2026 and divide the world into two roles.
A provider of a system that generates synthetic audio, image, video or text must mark the output in a machine-readable format so that it can be detected as artificially created. And whoever deploys and publishes such content, the deployer, must disclose that deepfake content has been artificially created or manipulated.
Creators and brands fall into the second category. A marketing department producing a product video with an AI avatar is a deployer. A creative studio that builds a campaign around a synthetic voice of a well-known face is a deployer. The obligation to label content rests with them, not with the tool that generated it.
Two dates instead of one
Entry into force rests on two dates. Most of the obligations under Article 50, including that a chatbot must identify itself as AI and that a deepfake must carry a label, apply from 2 August 2026 without delay. Machine-readable marking on the provider's side received an exception.
According to the text of the regulation and its interpretation, the grace period applies only to systems placed on the market before 2 August 2026, and only to the obligation to mark and detect AI content under Article 50(2). Providers of such systems must meet that obligation only from 2 December 2026.
The four-month window was created as part of the Digital Omnibus, a package of amendments that postponed some technically demanding obligations, but left disclosure towards the viewer unchanged from August.
For creators, this makes a practical difference. A deepfake that they publish in August must carry a visible label immediately. The machine-readable trace that the tool is supposed to embed in the file may not be visible until December with an older tool.
Sanctions and how they are interpreted
Non-compliance has a price. Fines reach up to €15 million or 3% of global annual turnover, whichever is higher. For EU institutions and bodies, the cap is set at €750,000.
Small and medium-sized companies calculate differently. The European Commission states in its FAQ that for SMEs and small mid-cap companies, proportionality can be taken into account, and the lower of the two values, percentage or fixed amount, applies to them. A start-up with small turnover therefore does not automatically face the fifteen-million cap.
How exactly national supervisory authorities will set sanctions remains to be seen. The concrete impact on individual creators and brands will depend on interpretation and enforcement, which will take shape in the first months of application. The Commission has also published a Code of Practice on Transparency of AI-generated Content, a voluntary tool with which a company can document compliance with transparency obligations, from labelling content to disclosing the artificial nature of images, audio and text.
AI video tools and the machine-readable trace
The new layer applies directly to the tools European creators work with every day. Synthesia offers over 240 AI avatars and over 160 languages and counts over 90% of companies on the Fortune 100 list among its clients. HeyGen advertises over 175 languages and dialects for translation and lip-sync, and for every custom avatar it requires identity verification of the person shown, directly on camera. Both companies are providers within the meaning of Article 50, and the machine-readable marking is on their side.
Proving the origin of content is not trivial, however. The standard for content credentials is C2PA, which writes metadata into a file about what it is and when it was created. According to the Fortune report, Synthesia does not label its videos with C2PA metadata by default, even though it is a member of the Partnership on AI and has completed a C2PA pilot.
Synthesia's head of corporate affairs argued in that report that “In the future, we are moving to a world where if something doesn't have content credentials, by default you shouldn't trust it.” The obstacle remains that, according to the same source, C2PA is computationally expensive and its metadata significantly increases file size.
This is where compliance becomes a business signal. A tool that can reliably embed a machine-readable trace makes the European deployer's own obligation easier. A brand that documents the origin of its AI video builds trust at a time when the share of synthetic content in feeds is rising.
The difference between providers is not only in the number of languages, however. Identity verification of the person shown on camera, which HeyGen requires for custom avatars, addresses a different part of the same problem, namely the consent of the person to be portrayed by AI. Article 50 requires that artificial origin be disclosed to the viewer, while standards such as C2PA address verifiability across files and tools. A European creator who assembles a campaign from several AI services therefore faces the question of which of them provide the compliance layer and which leave it to them.
Reach across borders
The rule does not stop at Europe's borders. A company based anywhere in the world falls under it if its content reaches a European audience. A creator who described the rule's effect under Digita's account on X summed it up as “posting AI-generated video that reaches EU viewers is covered.” In the same thread they added that content published before the rule took effect does not require retroactive labelling, but “anything new, going forward, does.”
The account nbtb pointed out a detail that gets lost in the discussion. Intent does not matter. Content that looks or sounds like a real person needs a label even when no one meant to deceive, and even when the person shown does not exist at all. A clearly creative, satirical or fictional work has, according to them, a milder disclosure regime, not a complete exemption.
For a European brand director, 2 August is the date from which AI video is also a question of compliance. The label on an avatar has become a legal obligation.
Comments