Labelling Obligations for AI-Generated Content – Implications for Advertising and Marketing

The use of AI has long become part of everyday practice for advertisers and marketing agencies. Content can now be created faster and more cost-effectively than ever before. However, these new opportunities are accompanied by increasing legal requirements.

While some of the deadlines for the application of the provisions of the AI Act (Regulation (EU) 2024/1689, the “AI Act”) have been extended by the so-called AI Omnibus, this is unlikely to apply to the labelling obligation under Article 50(4) AI Act. Accordingly, this obligation is expected to apply as early as 2 August 2026. In the fields of advertising and marketing in particular, the obligation to label certain AI-generated content – such as images and videos – is of significant practical importance.

This article outlines the specific requirements arising from these rules and examines the practical implications of the labelling obligation, particularly for advertising and marketing.

I. Key Legal Provisions

The central provision governing the labelling obligation is Article 50(4) AI Act. It requires operators of AI systems that generate or manipulate image, audio or video content qualifying as deepfakes to disclose that such content has been artificially generated or altered.

A corresponding obligation also applies to operators of AI systems that generate or manipulate text, provided such text is published in order to inform the public about matters of public interest. In such cases, it must likewise be indicated that the content has been artificially generated or modified.

Article 50(4) AI Act therefore establishes two distinct regulatory regimes: one for image, audio and video content, and another for certain types of text.

1. Images and Videos

The most relevant obligation in practice for companies in advertising and marketing concerns the labelling of AI-generated or AI-manipulated images and videos, provided these qualify as deepfakes. In this area, the AI Act imposes particularly extensive transparency obligations.

a) What qualifies as a deepfake?

A deepfake is defined in Article 3(60) AI Act 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.

There is currently significant legal uncertainty regarding the interpretation of this definition, which will likely only be clarified through future case law. A literal reading suggests that the content must resemble a real, specifically identifiable person or existing object, place, etc.

However, the prevailing view goes further: Even an abstract resemblance to fictional persons or objects may suffice, provided the content could be perceived as authentic or truthful. This approach is also reflected in the European Commission’s draft guidelines on the implementation of transparency obligations under Article 50 AI Act. According to these guidelines, it is sufficient that simulated persons, objects, places or events resemble something that exists – or could plausibly have existed – in reality.

By contrast, obviously unrealistic depictions are not covered. This includes content that contradicts the laws of nature or well-established biological principles, such as humans flying without technical assistance or dragons driving cars. The guidelines cite, for example, an AI-generated video of a mouse speaking with a human voice as content that would not qualify as a deepfake.

As a result, virtually any photorealistic depiction of human figures or realistic-looking objects, places or situations may qualify as a deepfake.

b) Classification of standard AI-assisted edits

It is controversial whether AI-assisted editing of existing content – such as adjustments to background details, lighting, audio parameters, colours or noise reduction – already leads to a classification as a deepfake and thus triggers a labelling obligation.

The view currently prevailing in many cases is that this is not necessarily the case, particularly where the respective edits play only a minor role in assessing the authenticity or truthfulness of the content. However, this always requires a case-by-case assessment.

The European Commission’s draft guidelines also differentiate in this respect. Substantial AI-based alterations – for example, changes to background details in journalistic images – may negatively affect the authenticity of the content. By contrast, measures such as colour correction, background extension of existing content or simple scaling of images (e. g. in product advertising) generally have only a limited impact on perceived authenticity.
A definitive clarification of these distinctions is still pending and will likely be left to future case law.

c) No privileged treatment for commercial advertising

Although Article 50(4), third sentence, AI Act provides for reduced transparency obligations for deepfakes forming part of clearly artistic, creative, satirical or fictional works or programmes, this exception does not apply to traditional product advertising according to the European Commission’s draft guidelines.

Companies therefore cannot generally rely on the argument that their advertising is merely creative or artistic in order to benefit from reduced transparency requirements. Moreover, the overarching objective of the AI Act – to prevent deception and ensure a high level of transparency – supports a broad interpretation of the labelling obligations.

d) AI models

The use of so-called “AI models” – AI-generated personalities that appear on social media as influencers or in advertising campaigns like real people – is particularly sensitive.

As such AI models are deliberately designed to resemble human beings and may even convey the impression of having a personal life, there are strong arguments that they qualify as deepfakes subject to labelling. Advertising measures involving AI models will therefore, in principle, have to include a corresponding indication of their AI-based origin once the labelling obligation becomes applicable.

2. Texts

The legal position for AI-generated text is significantly less strict. Article 50(4), second subparagraph, AI Act applies only to text intended to inform the public about matters of public interest. Traditional advertising copy, product descriptions, social media posts or newsletters will therefore generally fall outside its scope.

In addition, there is an important exception: No labelling is required where the text has been subject to human review or editorial control and where a natural or legal person assumes responsibility for it. This will typically be the case in marketing contexts. If AI-generated text is reviewed, revised and approved, a labelling obligation will usually not arise.

Nevertheless, it is advisable to implement clear internal processes and to document editorial responsibility transparently. Unreviewed AI drafts may not only entail legal risks but can also result in considerable reputational damage.

II. Who is subject to the labelling obligation?

According to its wording, the labelling obligation under Article 50(4) AI Act is addressed exclusively to the “operator” of an AI system. Under Article 3(4) AI Act, an operator is any natural or legal person who uses an AI system under their own responsibility, unless such use is purely private.

In practice, this means that essentially anyone who uses generative AI independently for commercial purposes will qualify as an operator.

However, it remains unclear how to assess cases where an advertising company does not use the AI system itself but merely uses AI-generated output created by an external agency. While operators must be distinguished from end users who simply use AI output without exercising control over the system, legal literature suggests that an “organisational” attribution of operator status may be possible. This could apply, for example, where a party assumes responsibility for the accuracy or completeness of the AI-generated output and third parties may reasonably rely on such responsibility.

Against this background, and in light of the protective purpose of the AI Act, there are strong arguments that the advertising company itself may also be subject to the labelling obligation – even where it did not create the AI-generated or manipulated content. Otherwise, such obligations could easily be circumvented by outsourcing advertising activities. Further clarification by the courts will be required in this regard.

III. Form of the labelling

Under Article 50(5) AI Act, the labelling must be provided in a “clear and unambiguous manner” and must comply with applicable accessibility requirements. The legislation does not prescribe a specific format.

Depending on the medium, target audience and context, possible forms include textual notices such as “AI-generated” or “created using AI”, visual indicators (e. g. pictograms) or auditory notices in the case of audio content.

Importantly, where content is directed at the German market, a German-language label should be used as a precaution. Purely English-language notices or hashtags such as #AIgenerated may carry increased legal risk, particularly in light of case law on influencer advertising disclosures.
IV. Potential consequences of non-compliance

1. Fines under the AI Act

Infringements of the labelling obligations may result in substantial administrative fines. Article 99(4)(g) AI Act provides for fines of up to EUR 15 million or up to 3 % of the total worldwide annual turnover, whichever is higher. While reduced maximum thresholds apply to small and medium-sized enterprises (SMEs), even these may represent a significant financial burden.

2. Unfair Competition law risks

From a practical perspective, the competition law dimension is particularly relevant. The labelling obligations under Article 50 AI Act are likely to qualify as rules governing market conduct within the meaning of Section 3a of the German Unfair Competition Act (UWG). As a result, a breach of the labelling obligation may also constitute an act of unfair competition.

This means that enforcement is not limited to regulatory authorities: Competitors, trade associations and consumer protection organisations may also take action. Warning letters, cease-and-desist claims and, in some cases, damages claims are therefore realistic scenarios.

V. Conclusion and practical recommendations

The labelling obligations under the AI Act are likely to have a tangible impact on the use of AI in marketing. This is particularly true for images and videos, given the broad interpretation of the concept of deepfakes and the tendency towards a wide application of the rules. By contrast, purely advertising texts are often less affected, as they frequently fall outside the scope of the provision or benefit from the editorial control exception.

Companies should therefore prepare for the new requirements without delay, especially since the rules will apply from 2 August 2026 and will also affect ongoing marketing activities.

In practice, it is advisable to establish clear internal guidelines for the use of AI in marketing. Where in doubt, providing a transparent indication (e. g. “AI-generated”) will generally be the safer approach. Such labelling should be clear, easily perceptible and – where content is aimed at the German market – generally provided in German.

This article was prepared with the support of our research assistant Franziska Klinzing.