Many forms of media created or significantly altered using artificial intelligence (AI) must be labelled as such in accordance with the AI Act. We explain when these AI transparency requirements specifically apply, what content must be labelled and how, and which exemptions are relevant to marketing.
Please note: The transparency requirements discussed here under Art. 50 AI Act will apply from 2 August 2026 and will not be postponed by the agreed amendments to the AI Act.
Why is the AI Act relevant to marketing?
The AI Act classifies AI systems into risk categories and links obligations to their respective role and context of use. For marketing departments, the ‘limited risk’ category and the role of the ‘deployer are particularly relevant (with one exception): Here, the focus is on transparency obligations – particularly in the case of synthetic or manipulated media, i.e. content that has been created or altered using AI. When using AI systems designed for interaction, such as chatbots on websites, the role of the ‘provider’ may also become relevant in the marketing sector.
Transparency obligations are intended to prevent people from unintentionally mistaking AI for human communication or artificial content for authentic recordings. The purpose of the Act is to reduce the risks of identity theft, deception, misinformation, manipulation and fraud.
In practical terms, transparency obligations under the EU AI Act mean that anyone using AI must make this clearly recognisable to those affected.
Which typical marketing use cases might trigger transparency obligations?
In marketing practice, AI transparency issues arise primarily in three scenarios:
- The publication of synthetic or manipulated media (image/audio/video), in particular featuring photorealistic-looking people or authentic-sounding voiceovers (so-called ‘deepfakes’);
- The publication of synthetic or manipulated text (e.g. LinkedIn posts, blogs, web articles), where this is not subject to editorial control;
- Human-machine interactions, such as chatbots or voice assistants on websites, whereby the company offering these services is deemed a provider within the meaning of the AI Act.
For marketing teams, this means that it must be made clear at the point of first interaction or perception that the content in question is AI-generated text without human review or editorial control, AI-generated or AI-manipulated content, or an AI chatbot.
What does the AI Act stipulate regarding media and deepfakes?
Art. 50(4) AI Act is particularly relevant for marketing using AI-generated or AI-modified images, audio or video files:
“Deployers of an AI system that generates or manipulates image, audio or video content constituting a deep fake, shall disclose that the content has been artificially generated or manipulated […] Where the content forms part of an evidently artistic, creative, satirical, fictional or analogous work or programme, the transparency obligations set out in this paragraph are limited to disclosure of the existence of such generated or manipulated content in an appropriate manner that does not hamper the display or enjoyment of the work.”
Art. 3(60) AI Act, in turn, defines the deepfakes referred to 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”
Thus, as soon as content appears authentic, the risk of misleading the public increases. The degree of realism and the potential for deception are therefore key criteria for identifying a deepfake, which in turn would trigger AI transparency obligations.
Two aspects are important here:
- The resemblance should be objectively recognisable.
- For the purposes of reality, it is sufficient if simulated persons, objects, places, facilities or events appear realistic and resemble something that could exist or may have existed in reality:
- The term ‘persons’ refers to realistic, natural human beings.
- ‘Objects’ refers to realistic, inanimate material objects, including buildings, works of art, machines, consumer goods, etc.
- ‘Places’ refers to realistic settings, such as a company canteen or a company’s headquarters.
- ‘Entities’ are to be understood as realistic, non-human but animate beings, including animals or other biological life forms.
- ‘Events’ are realistic scenes or situations in which people, objects, places and entities may be involved (such as historical events or the depiction of services).
Note: The transparency requirement also applies if the content recognisably imitates a person who once existed (a historical figure) and may therefore appear genuine. The purpose of protection (risk of deception or misrepresentation) applies equally to deceased persons.
When is there a risk that a person might mistakenly regard content as authentic?
When determining whether a deepfake is present, it is irrelevant whether the company or the marketing department intends to deceive anyone. What is decisive is how easily people could be deceived. In this regard, one should take the actual audience into account – that is, different age groups and levels of knowledge – rather than simply assuming the average characteristics of a target group. This is particularly important if the content created or altered using AI could also reach children, older people or those with limited digital or AI literacy, as they are more likely to misjudge its authenticity.
Assessing whether there is a risk that the content could be mistakenly perceived by a person as genuine or truthful is, by its very nature, a case-by-case matter and must take the context of each individual case into account.
Which AI-generated media do not require labelling?
Existing image, audio or video content that has undergone minor, technical editing using AI does not need to be labelled, as such AI-assisted editing of minor, technical aspects is of secondary importance when assessing the authenticity or accuracy of this content. This includes, for example:
- the editing of background details,
- lighting adjustments,
- the adjustment of audio parameters,
- colour corrections,
- noise reduction,
- improving accessibility, or
- file compression.
What constitutes works or programmes that are clearly artistic, creative, satirical, fictional or analogue?
To qualify for more lenient treatment, deepfake content must fall within one of the following five content categories: artistic, creative, satirical, fictional or analogue works or programmes. Whether deepfake content falls under one or more of these categories must be assessed on a case-by-case basis. Furthermore, the condition that content must “evidently” fall within one of the five content categories excludes content from the scope of these categories if it primarily serves an informative or commercial purpose and is recognisable as such.
In a marketing context, this minor exception is therefore not relevant to the transparency obligations, as marketing content generally serves an informative and/or commercial purpose.
Must AI-generated or AI-manipulated texts be labelled?
With regard to the publication of texts created using AI, Art. 50(4) AI Act stipulates:
“Deployers of an AI system that generates or manipulates text which is published with the purpose of informing the public on matters of public interest shall disclose that the text has been artificially generated or manipulated. This obligation shall not apply […] where the AI-generated content has undergone a process of human review or editorial control and where a natural or legal person holds editorial responsibility for the publication of the content.”
In marketing, therefore, a practical distinction applies to texts:
- If content has been created or modified using AI but is subject to human review or editorial control prior to publication, and a natural or legal person bears editorial responsibility for the publication of the content, it does not need to be labelled.
- If texts are to be published without human review or editorial control and no natural or legal person will assume editorial responsibility for their publication, the following criteria must still be met before a labelling requirement applies:
- The texts must be made available simultaneously and/or sequentially to an indeterminate, relatively large number of potential readers who are not connected to one another, regardless of whether this is done for a fee (e.g. as part of a subscription) or not.
- The texts must be intended to inform the public. This means that the texts should be designed to convey knowledge, opinions or facts.
- The texts must inform the public about matters of public interest. These are matters that are relevant to society (locally, nationally, at European level or internationally) and deserve public discussion or scrutiny. These include, for example, developments with potentially significant implications for the economy, politics, science or culture. The examples are not exhaustive; what constitutes public interest may vary depending on the context and the time.
Note: Texts that are not considered to be published include, for example, private, personal correspondence (for professional purposes) or internal organisational documents.
As is clear from the above, it will only rarely be necessary for the marketing department to label a text.
When must AI chatbots be labelled?
Art. 50(1) AI Act applies to AI-based human-machine interactions:
“Providers shall ensure that AI systems intended to interact directly with natural persons are designed and developed in such a way that the natural persons concerned are informed that they are interacting with an AI system, unless this is obvious from the point of view of a natural person who is reasonably well-informed, observant and circumspect, taking into account the circumstances and the context of use.”
In a marketing context, this becomes relevant when, for example, chatbots or dialogue agents are integrated into e-commerce platforms or websites. Anyone requesting the integration of such chatbots from the IT department must be aware that the company may thereby be regarded as a ‘provider’ and must therefore comply with the transparency obligations when developing the chatbot or commissioning a third party to develop it.
In practice, however, the marketing department will be able to do little – or indeed be required to do little – in this regard.
What is meant by ‘first interaction’ or ‘first contact’?
Art. 50(5) AI Act requires that transparency obligations be fulfilled no later than the time of the first interaction or exposure. The ‘first interaction or exposure’ refers not only to the first natural person who interacts with an AI system (or its output, such as an AI-generated video) or is confronted with it, but also to any subsequent first interaction with the AI system (or its output) or any subsequent first exposure to it by another person. Therefore, it is not only the first viewers who count, but also all subsequent potential viewers.
As people may, for example, only begin watching once a live broadcast featuring deepfakes has started, the warning should not only be given at the start of the broadcast, but also at later points or on an ongoing basis. In the case of published deepfake content, it must be ensured that the labelling remains visible to all subsequent viewers, even if the deepfake is republished elsewhere.
What form should the AI label take in practice, and where should it be placed?
Publishers of deepfake content or AI-generated or AI-manipulated text must, in accordance with Recital 134 of the EU AI Act, ‘clearly and unambiguously disclose that the content has been artificially generated or manipulated by labelling the AI output accordingly and disclosing its artificial origin’.
The European Commission’s draft ‘Guidelines on the implementation of transparency obligations for certain AI systems pursuant to Art. 50 of the AI Act’ makes it clear that the labelling must be affixed ‘on the content’ – not hidden in metadata or merely in a caption, which can easily be cropped out.
No specific disclosure method or format is prescribed for meeting the transparency obligations for AI chatbots. Examples of how the information might be provided include written notices (e.g. a chatbot that says ‘You are currently interacting with an AI system’ at the start of a conversation), visual means (e.g. an email generated by an AI agent to a natural person, bearing an AI notice at the top, or permanently displayed badges that remain visible throughout the interaction) or audio cues (e.g. a voice assistant that says ‘This is an AI-powered assistant’ at the start of a session, combined where appropriate with regular reminders during longer interactions).
Accessibility is also important where relevant: according to the AI Act, labels must be designed in such a way that they can be reliably perceived by people with disabilities.
Tip: Please also read our explanation on website accessibility in accordance with the German Accessibility Improvement Act (BFSG) and any necessary accessibility statement.
How can marketing teams implement AI transparency obligations in practice?
Transparency obligations are met in day-to-day operations through processes – not through one-off training sessions. The following building blocks are particularly effective for marketing departments:
- Develop standard texts and label variants for different channels (website, social media, video, audio).
- Ensure assets are tagged in your file directories, e.g. ‘AI-generated’, ‘AI-modified’, ‘Human-reviewed’.
- Pre-publication review process: a brief mandatory check before publication (including a deepfake check) and editorial review of texts.
What happens to content that has already been created and published?
AI-generated or manipulated content (including deepfakes) created before 2 August 2026 and already made available does not need to be retrospectively labelled or marked.
However, in view of the AI Act’s objective of promoting transparency and strengthening trust and integrity within the information ecosystem, the European Commission’s draft guidelines encourage deployers of AI systems and other stakeholders who possess and/or disseminate such content to do so.
What penalties apply for breaches of transparency obligations?
The AI Act provides for fines of up to EUR15,000,000 for breaches of transparency obligations or – in the case of companies – up to 3 per cent of the global annual turnover of the preceding financial year, whichever is higher.
In addition, there may be further consequences under other areas of law, such as the GDPR (in the case of unlawful data processing) or consumer and fair trading law (in the case of misleading advertising).
Conclusion: Transparency is not a label – it is a process
If you reduce transparency obligations to an AI watermark, you will fail in your day-to-day operations. Marketing departments that
- establish simple decision-making criteria for deepfakes and AI-generated text,
- standardise labelling appropriate to the channel and format,
- make approval and documentation procedures streamlined yet binding, and
- organise training on the new requirements.
Here’s how to meet the transparency requirements of the AI Act in a practical way – whilst reducing the risk of misleading communication.
