Who labels AI content, and when
On 2 December the EU’s marking duty reaches older AI tools, but the company that publishes a deepfake has had to label it since August.
Many companies have 2 December in the calendar as the day AI content must be labelled in the EU. That is only half right. The date belongs to the makers of generative tools, and to one duty: marking what their systems produce so that machines can detect it. The company that publishes a deepfake has had its own duty since 2 August, and the extra months do not apply to it.
Here is who carries which duty, what counts as a deepfake, where the exceptions lie and why a Swiss company is in scope. It is an overview, not legal advice.
Who marks, who labels
Article 50 of the EU AI Act splits the work in two. The provider, the company that develops a generative system and offers it under its own name, must ensure that its outputs are “marked in a machine-readable format and detectable as artificially generated or manipulated.” The deployer, the company that uses the system in its work, must disclose a deepfake: it “shall disclose that the content has been artificially generated or manipulated.” A company that builds its own generator on top of a model and puts it into service under its own name can become a provider as well.
Personal, non professional use falls outside the act, but a person who earns from it on a regular basis counts as a deployer. A company remains the deployer when contractors or freelancers run the system on its behalf and under its control.
The Digital Omnibus on AI, Regulation (EU) 2026/1744, published on 24 July, gave one group more time. Generative systems placed on the market before 2 August 2026 must meet the marking duty in Article 50(2) by 2 December 2026. Systems launched since then have had to mark from the start. The Commission says the grace period applies “only as regards the marking and detection obligation”, so nothing moved for the deployer. Content made before 2 August does not need a label after the fact.
The mark in the file does not replace the label. In the Commission’s words, deployers “cannot simply rely on the machine-readable marking embedded in the content by the provider”. People have to be told in a clear and distinguishable way, at the latest when they first see or hear the content.
What counts as a deepfake
The act defines a deepfake as AI generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and “would falsely appear to a person to be authentic or truthful.” The Commission reads three criteria into it: high resemblance, a subject that exists or “can plausibly exist”, and the capacity to mislead. A photorealistic person who never lived can count, and so can a cloned voice, because audio is covered as much as image and video.
The Commission’s final guidelines, published on 20 July, add that content may fall outside the definition where the audience does not expect it to be authentic. Their examples are less forgiving for marketing. According to a summary by the law firm Faegre Drinker, generated images that make products appear different from reality, digital replicas of real people and actors made to look younger all count as deepfakes that need disclosure. Text is never a deepfake. It has a rule of its own.
Art, satire and edited text
Where a deepfake is part of an “evidently artistic, creative, satirical, fictional or analogous work or programme”, the duty shrinks to disclosing it “in an appropriate manner that does not hamper the display or enjoyment of the work.” The word evidently carries weight. Faegre Drinker reads the final guidelines as accepting advertising as creative only in narrow circumstances, with most advertising examples outside the lighter regime.
For text, the duty covers writing published to inform the public on matters of public interest. The Commission lists politics, public health, consumer safety and any economic, financial, political, scientific or cultural development that may be a subject of public debate, so the rule can reach a company’s blog or newsletter as well as a newsroom. It falls away if the text has undergone human review or editorial control and someone holds editorial responsibility. The bar is real: spell checking and grammatical correction do not count, and someone must hold “the ultimate legal responsibility” over the publication.
Why Swiss companies are in scope
The act reaches beyond the EU. It applies to providers and deployers in third countries “where the output produced by the AI system is used in the Union”. A Swiss company that runs a campaign with generated faces for buyers in Germany, or a video with a cloned voice for customers in France, should treat the labelling duty as its own. National market surveillance authorities enforce it, and the Commission puts the fines for breaches of the transparency duties at up to 15 million euros or 3 percent of worldwide turnover.
At home there is no AI law yet. The Federal Department of Justice and Police is to present a consultation draft by the end of 2026, with transparency among its topics. Until then the data protection act applies, and the FDPIC says the use of programs that manipulate faces, images or voice messages of identifiable persons “must also always be clearly indicated.”
What to do before December
List the generative tools in use, including the ones staff found for themselves, and ask each provider in writing how it marks output and from when.
Check that your own workflow keeps those marks. Exports, compression and editing can strip metadata. The Code of Practice on transparency of AI generated content, published on 10 June, has providers layer their marks, generally signed metadata plus an imperceptible watermark, because no single technique is enough.
Decide how you label. The Commission and the AI Board have confirmed the code as an adequate voluntary tool, and the EU offers a set of icons for the purpose. Agree who places the label in each channel, agencies included.
For text, name the editor who signs off pieces on matters of public interest. Keep the prompts, the originals and the decisions with every published piece. If an authority asks, the record is the answer.
Where we stand
Reel, which is live, makes images, film and 3D with AI, and a person picks every image and clip that is used. Each one sits in a media library with every version and the recipe behind it. Studio, in early access, keeps a company’s marketing in one place, where a person approves every change before it goes live and every action lands in an audit log. Neither replaces the label. Both make it easier to know where one belongs.
