If you publish AI-generated video, images, audio, or public-interest text to people in the European Union, you have until 2 August 2026 to add clear disclosure. The rule splits in two: the company that built your generator embeds a machine-readable signal, and you, the publisher, add a label a person can actually see or hear. Miss it and national regulators can issue fines of up to €15 million or 3% of global annual turnover, whichever is higher.
This guide covers what to add to each format and gives you a pre-publish checklist you can reuse on every asset. For the legal background and who drafted the rules, see our explainer on the EU's AI content labelling Code of Practice.
Who actually has to label AI content?
The AI Act separates two roles. Providers are the tool makers, such as the company behind a video generator, and they embed a hidden provenance signal into every output. Deployers are the people and brands who publish that output. If you post an AI clip to your channel or hand an AI voiceover to a client, you act as a deployer, and the visible disclosure becomes your responsibility.
Company size gives no exemption. The trigger is the audience: content that reaches EU users falls in scope whether you are a solo creator or an agency. A freelancer producing AI product ads for a German shop and a marketing team localizing an AI explainer into French both count as deployers, and both owe a label.
Two categories always need disclosure. A realistic deepfake, meaning AI video or audio that depicts a real person or event, needs a clear label. AI-generated or AI-edited text published to inform the public on matters of public interest needs one too, with a single exemption: text that a named human editor has reviewed and taken responsibility for. Matters of public interest cover reporting on news, elections, public health and comparable civic topics, which is where an unlabeled AI edit does the most damage. Conversational AI sits on its own line, where you tell users they are dealing with a machine before the chat begins.
What do you add to an AI video, image, or audio file?
The Code of Practice sets one principle above the rest: a person must notice the disclosure without special tools. A hidden watermark answers the provider's duty, not yours. Placement then shifts by format, because a badge that works on a still image does nothing for a podcast. The Code also standardized the wording into two short icons, shown as "AI" in English, "KI" in German, and "IA" in French, with equivalents for other member-state languages.
Here is the working version for each format:
- Video: keep a persistent on-screen label for the full length of the clip, not a single flash at the start. Apply the approved "AI" mark and localize its text for the market you target. Leave the generator's embedded metadata intact when you export.
- Images: put a visible marking on the image itself or in the caption sitting right next to it, and keep any C2PA or provenance data the tool wrote into the file.
- Audio and voice: record an audible spoken disclaimer at the opening, and repeat it after long segments, since a listener has nothing to look at.
- Public-interest text: add a visible note that the piece is AI-generated or AI-assisted, unless a named human editor signed off on it.
- Interactive AI: state that the user is dealing with an AI system before they send a first message.
The two approved icons carry no attribution requirement, so you can place the "AI" mark yourself without licensing it or crediting a source. Live and streamed content follows the same logic, with the label held on screen or repeated aloud for the duration of the broadcast.
The pre-publish labelling checklist
Run this quick pass before anything AI-made goes live to an EU audience:
- Decide whether the piece is a realistic deepfake or public-interest text. If it is, treat disclosure as mandatory rather than optional.
- Add the human-visible or audible label in the correct spot for the format.
- Confirm the provider's machine-readable signal survived your edit and export, because re-encoding often erases it.
- Switch on the platform's own synthetic-content toggle where one exists, as a second layer.
- Localize the icon text for each EU market you target.
- Save the labeled version as a template so your next upload starts compliant.
Our breakdown of the platform disclosure settings on YouTube, TikTok, and Instagram shows where each toggle sits and how it maps to this rule.
What are the penalties, and when do the rules start?
Two dates shape this summer. By 22 July 2026, providers and deployers who want the firmest legal footing submit their signatory forms to the EU AI Office to appear on the first published list and secure a presumption of conformity. From 2 August 2026, the Article 50 transparency obligations apply across all 27 member states, and that is the date publishers feel directly.
Enforcement rests with national market surveillance authorities, so the office that contacts you depends on the country where the content lands. The ceiling reaches €15 million or 3% of worldwide annual turnover, whichever is greater. There is a break for smaller operators: for SMEs and startups, the Act flips the formula and caps a fine at the lower of the fixed amount or the percentage, so a solo creator faces a proportionate figure rather than the headline €15 million, though the duty to label stays identical.
The Code itself stays voluntary, yet signing it grants a presumption that you comply, which shifts the burden if a regulator ever asks. Regulators publish the initial signatory list before 2 August, so a late signature still leaves you off that first roster. You can read the duty itself in Article 50 of the AI Act and the marking detail in the Commission's Code of Practice.
A common mistake that quietly voids your disclosure
The most frequent error in draft compliance plans is treating a hidden signal as the entire answer. A team switches on SynthID-style watermarking or the platform's back-end "altered content" flag, then assumes the file is handled. It is not. Those signals satisfy the provider side and feed detection tools, while the AI Act wants something a viewer or listener perceives on their own. Skip the visible or audible layer and a realistic deepfake stays unlabeled in the eyes of the rule.
The export step hides a second trap. Provenance metadata often disappears when a clip is re-rendered or compressed for upload, which means the signal you trusted never reaches the audience. A persistent corner label survives re-uploads and screen recordings far better than a caption a platform can crop, so pair a visible mark with the embedded data instead of choosing one. For the technical side of provenance, our guide to SynthID and provenance watermarking goes deeper.
The safest way to stay compliant
Build disclosure into your production template once, rather than deciding case by case under deadline pressure. Using the checklist above, set your format defaults, then verify the embedded signal after each export. Handled that way, a label adds a few seconds per asset instead of a legal review per post, and your channel stays clear well before the 2 August 2026 cutoff.
