Labelling Deepfakes: What the AI Act Requires of Companies from 2 August 2026
(As of: June 2026)
From 2 August 2026, the transparency obligations of Art. 50 of the AI Act apply. Anyone using AI-generated image, audio or video content in marketing, corporate communications or on social media must, under certain conditions, disclose it as artificially generated or manipulated pursuant to Art. 50 of the AI Act. But what actually constitutes a "deepfake" within the meaning of the Regulation? And how exactly must a deepfake be labelled?
Both questions determine concrete obligations, the breach of which can even entail risks under competition and regulatory law.
What is a deepfake subject to labelling within the meaning of the AI Act?
The legal definition of the deepfake is found in Art. 3 No. 60 AI Act. The AI 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".
This yields four elements: The content is
(1) AI-generated or AI-manipulated,
(2) is used professionally,
(3) resembles real persons, objects, places or events, and
(4) would appear authentic to a person.
Condition (1): AI-generated or AI-manipulated
-
Generated: The content is newly created by the AI. There is no real source material; the model produces image, audio or video from scratch (for instance a fully AI-generated product photo, a synthetic voice, a text-to-image motif).
-
Manipulated: There is real, existing source material that the AI alters – face swaps, adding or removing objects, voice or facial expression changes, composite images.
When is there no labelling obligation, no deepfake?
Not every AI touch on source material turns the content into a deepfake subject to labelling. The actual legal threshold is the substantial alteration. What matters in the case of manipulation is that the source material is substantially altered. Only the semantic change – where meaning, style or intention are affected – turns the intervention into a deepfake subject to labelling.
Guidance is provided by the draft Commission Guidelines on Art. 50 AI Act of 8 May 2026.
Important: These guidelines are not yet final – the consultation phase ended on 3 June 2026 and the final version is still pending. They are not legally binding but reflect the Commission's authoritative line of interpretation. The paragraph numbers below refer to the draft version.
According to it, an alteration is substantial and therefore subject to labelling if input data or its semantics were significantly altered by the AI system in the generation process; this is assessed by format, media type and style, and by whether the alteration affects the meaning, style or intention of the content (para. 85 of the draft). It always remains a case-by-case assessment, and this standard applies in parallel to providers and deployers within the meaning of the AI Act.
- Not covered is mere standard editing (para. 84): correction, format conversion, minor cropping, noise reduction, minor colour or brightness adjustments. Purely technical minimal interventions – background details, lighting, audio parameters, compression – generally do not turn content into a deepfake either (para. 109).
- Covered, by contrast, is the semantic change (para. 86): AI translations and summaries, adding or removing objects, pixelating faces, altering body shape or skin colour, extreme colour and contrast interventions, the conversion of black-and-white to colour, and composite images.
Purely technical minimal interventions, such as minimal post-processing (cropping, colour correction, noise reduction), can therefore remain out of scope. With generated content this question does not arise, because the content was artificially created in the first place.
The standard is, moreover, context-dependent and must be applied more strictly, for instance, to journalistic imagery. The rule of thumb: Whoever changes meaning, style or intention must label; whoever merely makes technical improvements need not.
Condition (2): The content is used professionally
Deepfakes only become subject to labelling insofar as they are used in a professional context, such as advertising and brand presence.
This follows from Art. 2(10) of the AI Act, which excludes the labelling obligation for use in a purely personal context.
Condition (3): resembles real persons, objects, places or events
Deepfakes are not limited to persons. Objects, places, entities and events are expressly covered as well. The AI-generated product image or the fictitious location can be affected by the labelling obligation just as much as the manipulated portrait.
According to the Commission guidelines, it is sufficient if simulated persons, objects, places, entities or events resemble someone or something that exists or could have existed in reality.
Condition (4): appears "authentic"
In addition, what is depicted must appear "authentic".
The decisive benchmark is not the reasonable average viewer but the actual – often heterogeneous – composition of the audience; particular consideration must be given to vulnerable groups such as children or the elderly (Commission draft guidelines, para. 108). An intention to deceive is irrelevant.
The underlying rationale is consumer protection: In case of doubt, the addressee should be able to recognise that they are dealing with artificial material.
Who must label what? Providers and deployers
The AI Act distinguishes two categories of obligated parties – providers and deployers – with different requirements.
Providers
A provider within the meaning of Art. 3 No. 3 of the AI Act is:
"a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge"
The provider is subject to a machine-readable marking obligation: The outputs of the AI system must be marked in a machine-readable format and be detectable as artificially generated or manipulated (Art. 50(2) sentence 1 AI Act). The information should be technically embedded in the metadata.
For the element of machine readability, recourse can be had to the understanding of the term under Art. 20(1) GDPR: machine-readable formats are those that enable software- or computer-based processing.
Deployers
A deployer within the meaning of Art. 3 No. 4 AI Act is:
a natural or legal person, public authority, agency or other body using an AI system under its authority except where the AI system is used in the course of a personal non-professional activity.
In other words, companies that use a deepfake. Deployers must disclose that the content has been artificially generated or manipulated (Art. 50(4) sentence 1 AI Act). This disclosure is addressed to humans and must be perceptible to them. For most companies, this is the practically decisive obligation.
How must content be labelled?
The modalities of the labelling obligations are governed by Art. 50(5) AI Act:
The disclosure must be clear and distinguishable and made at the latest at the time of the first interaction, Art. 50(5) AI Act. From this and from the recitals, strict requirements follow:
- The labelling must be clear, distinct and easily perceptible. A blanket disclaimer, for instance in the legal notice, is not sufficient. What is required is a context-related, prominent notice.
- There is no room for creative paraphrasing. The notice must be stated clearly, for instance "AI-generated", "created with AI" or "synthetic content", alternatively as a visible watermark or label.
- The notice must be provided upon first perception. Providing information after the fact is therefore not sufficient.
For the "how", the Commission provides first illustrative examples in the Annex to the EU Code of Practice on Marking and Labelling of AI-generated Content (Art. 50 AI Act), 2nd draft of 5 March 2026 – the "Sample icons developed by the European Commission", the voluntary code currently at draft stage (final version expected ~June 2026).
You use AI-generated content in marketing or communications and are unsure what will be subject to labelling from August? Schedule a free initial assessment.
Exceptions: personal use, art and satire
In addition to the exception for the private use of AI-generated content under Art. 2(10) AI Act, Art. 50(4) subpara. 2 AI Act contains a relaxation for evidently artistic, creative, satirical or fictional works.
Contrary to what is occasionally assumed, the obligation does not lapse entirely here: It is reduced to a disclosure made in an appropriate manner that does not hamper the display or enjoyment of the work. Where exactly the line runs in the tension between the labelling obligation and artistic freedom remains open – the first court decisions will have to be awaited.
From when does this apply? Cut-off date, retroactivity – and the "Digital Omnibus"
The transparency obligations apply in principle from 2 August 2026. They do not apply retroactively: Existing content does not have to be labelled after the fact. Only content newly generated or manipulated from that cut-off date is covered.
At this point, however, a current development must be factored in that is missing from many accounts – and which must be sharply distinguished between the deployer and provider obligations:
- Deployer obligation (deepfake disclosure, Art. 50(4)): Here, 2 August 2026 stands. This date is not affected by the ongoing reform debate and is fixed. For most companies – and this is the practically important obligation – nothing changes in the timetable.
- Provider obligation (machine-readable marking, Art. 50(2)): Here there is movement. On 7 May 2026, Parliament and Council provisionally agreed on amendments as part of the "Digital Omnibus on AI". Under this, generative AI systems already placed on the market before 2 August 2026 are to benefit from a transitional period – with 2 December 2026 under discussion. However, this agreement is not yet applicable law; it requires formal adoption and publication in the Official Journal. If the procedure drags on beyond August 2026, the existing deadlines remain in place for the time being (cf. Art. 113 AI Act).
For planning purposes, therefore: The deepfake labelling in your own marketing should be firmly geared to 2 August 2026. Where you rely on machine-readable markings of your tools on the provider side, it is worth keeping an eye on the state of the reform – this technical layer may still shift in time. In parallel, the AI Office is working on a Code of Practice on the labelling of AI-generated content, the finalisation of which is expected for mid-2026; it will further substantiate practical implementation.
Missing labelling from 2 August – what legal consequences loom?
In the event of missing labelling within the meaning of the AI Act, authorities can intervene and there is a risk of fines.
The exact scope, in particular the significance of Art. 50 AI Act for competition law, has not yet been conclusively clarified in the absence of case law on Art. 50 AI Act.
The parallel to influencer advertising labelling suggests, however, that the courts will take a consistent line here. For companies with their own clean labelling practice, this opens up a solid position of attack – up to and including preliminary injunction proceedings where urgency exists.
You want to align your content processes with the new transparency obligations in a legally compliant way or take action against a competitor? We examine which of your content is subject to labelling, develop practicable labelling standards and support you in enforcement. Schedule a free initial assessment.
FAQ
1. Do I have to label every AI-generated image?
No. Only content that crosses the threshold of "substantial alteration" and resembles real persons, objects, places or events is subject to labelling. Mere standard editing – cropping, colour correction, noise reduction – is generally not sufficient for that.
2. Is a notice in the legal notice or the profile bio sufficient?
No. The disclosure must be clear, distinguishable and made upon first perception – i.e. visible on the content or immediately next to it, not hidden in a subordinate place.
3. Label in German or English?
If the content addresses the German market, German should be chosen in case of doubt. Unclear or purely technical English terms are risky in line with the case law on advertising labelling.
4. Does the obligation also apply to AI content I created before 2 August 2026?
No. The obligation does not apply retroactively. Only content newly generated or manipulated from the cut-off date is covered.
5. Will the cut-off date shift due to the "Digital Omnibus"?
Not for the deployer obligation to disclose deepfakes (Art. 50(4) AI Act) – here 2 August 2026 stands. Only the provider-side machine-readable marking (Art. 50(2) AI Act) may receive a transitional period for legacy systems; however, this has not yet been finally adopted.
6. What are the consequences of infringements?
In addition to regulatory fines (up to EUR 15 million or 3% of worldwide annual turnover), there is the threat of cease-and-desist letters under competition law and injunctive relief claims by competitors and associations.
