Mice that argue in human speech over the best cheese in a cheesemaker’s advert are not deepfakes, according to the European Commission’s guidelines. A realistic synthetic influencer trying out a sponsored product is. The AI tool embeds the machine-readable marking in the file, but a creator who regularly earns money from content must disclose a deepfake themselves.
A person in front of a screen cannot recognise a machine-readable marking without technical aids. From this, the Commission concludes in its guidelines on Article 50 that “deployers cannot rely on the machine-readable marking embedded in the content by the provider”. This also applies to the marking that ChatGPT, Gemini or Sora embed in their outputs, Notebookcheck points out. The obligation to disclose a deepfake applies from 2 August 2026 with no transitional period. Older AI tools received a postponement for the machine-readable marking.
It is precisely the machine-readable marking that, according to the Czech Association for Artificial Intelligence, causes the greatest confusion in the public debate about labelling AI content. In August, Lupa headlined its report on the recommendations of the Czech Telecommunication Office (ČTÚ) “You must now label AI creations”. The association, however, points out that the AI Act “does not say that every AI image, every AI text or every chatbot output must get a visible sticker”.
The tool provider is responsible for the marking in the file
Paragraph 2 of Article 50 requires providers of AI systems that generate synthetic audio, images, video or text to mark their outputs. According to the text of the regulation, these are to be “marked in a machine-readable format and detectable as artificially generated or manipulated”. The obligation covers all such content except for specified exemptions. The user of the tool does not have to embed a machine-readable marking.
An exemption applies to edits where AI merely assists with standard editing or does not substantially alter the input. The guidelines list noise removal, minor cropping, colour and contrast adjustments and limited video stabilisation among them. A brightened shot in a vlog does not need the marking.
How to embed the marking is described in the voluntary code of practice that the Commission published on 10 June 2026. According to it, providers should as a rule mark an output with at least two techniques, such as digitally signed metadata and an invisible watermark. By the end of July, according to the Commission, the code had been signed by roughly 190 organisations, Google among them.
The law firm Paul, Weiss summed up the code in a client memo as “Voluntary in name, important in practice.” According to the firm, anyone who has not signed it must defend to the supervisory authority on their own that their measures are sufficient.
The deepfake label is applied by the creator with a monetised channel
The Commission’s guidelines of 20 July 2026 explicitly count among professional activity also one from which a person has “an economic benefit on a regular basis”. Anyone who uses AI this way is a deployer (the Czech term used by ČTÚ is “zavádějící subjekty”). ČTÚ uses the term for persons who “use an AI system in activity other than personal non-professional activity”, and it is on them that the obligation to disclose a deepfake and AI text on matters of public interest falls.
Notebookcheck summed it up for creators with the sentence “The European Commission draws the line at money, not at reach.” According to it, a monetised YouTube channel and an Instagram account with paid collaborations usually suffice. The Czech legal guide Můj právník draws the same line. An AI image shared in a private conversation or among friends creates no obligation.
A deepfake is, according to the definition in the regulation, AI imagery, audio or video that “resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful”. Among its examples, the Commission’s guidelines name an AI portrayal of an influencer celebrity in an advert or the cloned voices of the regular hosts of a daily news podcast.
By contrast, a radio programme in which AI only evens out the volume and removes noise without changing the words is not a deepfake. According to the guidelines, the same applies to a real car in an advert set against an AI-generated background, provided the advert does not mislead about what the product looks like and what it is for.
The label must be visible or audible to a person without technical aids, at the latest at the first exposure to the content. According to ČTÚ, one can use, for example, the label “AI generated” for fully generated content and “AI modified” for edited content. Ready-made EU icons for both variants can be downloaded free of charge in SVG and PNG formats, and using them is voluntary, just like the code. For an audio deepfake, a notice should be heard, typically briefly at the start of the content. Information hidden in metadata or in terms and conditions is not enough.
For a clearly artistic, satirical or fictional work, the obligation is relaxed, and it is enough to make the disclosure in a way that does not spoil the experience of the work. The exemption for human review and editorial responsibility applies only to AI text that informs the public on matters of public interest, and it does not apply to deepfakes. AI advertising copy or a product description is not subject to this obligation if it contains no claims about, for example, health, consumer safety or sustainability. The synthetic influencer from the introduction meets neither of these conditions.
December ends the postponement for older AI tools
The postponement came with the Digital Omnibus, that is Regulation (EU) 2026/1744. Generative tools that were already on the market before Article 50 took effect must mark their outputs machine-readably from 2 December 2026. Newer tools must embed the marking from the moment they are placed on the market.
For a creator who merely uses an AI tool, nothing changes in December. A creator does not have to retroactively label a deepfake that was made and published before August, though the Commission encourages doing so where possible.
The supervisory authority in Czechia is to be, among others, ČTÚ. A breach of Article 50 carries a fine of up to EUR 15 million or 3% of worldwide annual turnover, whichever is higher. For small and medium-sized companies, the lower of the two amounts applies. As a first step, ČTÚ recommends correctly determining one’s own role.
From December, even the cheese mice must carry the machine-readable marking, if an older tool generates them only after that. They do not need a visible label even then. The synthetic influencer needs one from August, and it must be applied by the creator who regularly earns money from content.
Comments