From August 2026, the transparency requirements set out in Article 50 of the AI Regulation (‘AI Reg’ or ‘AI Act’) will apply. Now, more than ever, many companies are asking themselves: Do we need to label AI-generated content – such as website posts, newsletter texts or images? If so, when, how and to whom? And above all: What does this mean in concrete terms for our website, our chatbot or our product images?
1. What AI content must be labelled?
Article 50 of the AI Regulation sets out transparency and labelling requirements for AI systems that generate or manipulate content and whose outputs are perceived by humans – i.e. are visible, readable or audible. The reason for the labelling requirement is the high risk of manipulation or confusion posed by such AI systems – as their capabilities are very similar to those of humans, and we need to know when we are dealing with authentic content and when with synthetic content.
The labelling requirement therefore applies to AI systems that
- are intended for direct interaction with natural persons (e.g. chatbots), para. 1;
- are designed to generate synthetic content (audio, images, video or text), para. 2,
- are intended to generate deepfakes, para. 4, variant 1, or
- are intended to generate or manipulate text intended to inform the public, para. 4, variant 2,
or – though this is generally not relevant in most companies’ practice – if
- the AI system is an emotion recognition system or a system for biometric categorisation, para. 3.
Overview of AI systems subject to labelling requirements:
| Use case | Who is responsible? | Example |
|---|---|---|
| Interactive system (para. 1) | Provider | Chatbots, voicebots |
| Synthetically generated content (para. 2) | Provider | AI-generated product images |
| Deepfakes (para. 4, variant 1) | Operators | AI-generated images of a celebrity |
| Texts on topics of public interest (para. 4, var. 2) | Operator | AI-generated press release |
| Emotion recognition, biometric categorisation (para. 3) | Operator | AI for facial analysis |
Let’s take a closer look at the use cases:
Interactive AI systems
When people communicate directly with AI systems, they should be able to clearly recognise that they are not interacting with a human being. The duty of transparency serves to protect against deception or identity fraud.
Direct interaction occurs when the AI system is designed in such a way that people might think they are dealing with another real person. What matters, therefore, is not the technology behind the scenes, but the external impression.
This always concerns situations in which humans and machines engage with one another – that is, react to one another. This can take place in writing, verbally, or even through gestures or movements. What is crucial is the two-way communication between humans and AI.
Examples:
- Chatbots on websites or helplines that communicate via text or speech
- Voice assistants that respond to spoken commands
- Social robots, e.g. in care or industry
- Social bots that post, comment or like on social media and appear like real people
- Apps that, for example, act as human opponents in games, even though an AI is actually playing
Synthetically generated content
This refers to content that has not been created by a human but has been generated artificially and appears ‘real’ even though it is not. People should always be able to distinguish artificial content from genuine content in order to prevent deception or fraud, identity theft and the manipulation of opinions.
Examples:
- Articles, emails or stories written by text generators such as ChatGPT
- Images artificially created by image generators such as DALL·E or Midjourney using text inputs
- Videos in which people say or do things that never actually happened
- AI voices that mimic a human voice to a deceptively realistic degree (e.g. for fake calls)
- Avatars or virtual influencers that appear to be real people but are entirely AI-generated
Deepfakes
The law defines ‘deepfakes’ as AI-generated or manipulated image, audio or video content ‘that resembles real persons, objects, places, institutions or events and would falsely appear to a person to be genuine or truthful.
What are AI deepfakes?
≠ Deepfake: obviously fictional content
- Examples: comics, fantasy depictions, a unicorn galloping over a rainbow, a mermaid in a supermarket
- Not a deepfake = no labelling requirement
✓ Clear deepfake: existing people, places, events
- Depictions of real people, places or events, provided they are AI-generated or manipulated and capable of deceiving, e.g.:
- an AI-generated video falsely showing the Chancellor waving a rainbow flag on the balcony of the Chancellery,
- an audio recording in which a celebrity’s voice is artificially replicated to spread false statements,
- an image of a fictional demonstration in front of the Brandenburg Gate that never actually took place.
- Deepfake = labelling requirement
? Deepfakes: fictional but realistic-looking content
- It is disputed whether photorealistic – but in fact non-existent – fictional AI content must also be labelled as deepfakes, e.g.
- AI-generated, realistic-looking models presenting clothing in online shops,
- deceptively realistic so-called ‘AI models’ (virtual influencers) who present themselves as human-like online but are not modelled on real people,
- photorealistic, AI-generated people set against fictional but authentic-looking landscapes,
- podcast dialogues between non-existent people that sound to the listener like a conversation between real people.
Does fictional but realistic-looking content need to be labelled?
Yes. If you want to be on the safe side, you should also label AI-generated content that, although photorealistic, does not depict real people, objects or places.
This applies at least whenever the content is likely to deceive the average viewer and be taken for genuine, as is the case, for example, with deceptively realistic ‘AI models’ or authentic-looking videos of purely fictional events.
Check question:
- will the average viewer not assume that the person, object or place depicted is real -> labelling requirement (-)
- Will the average viewer assume that the situation depicted is not real? -> Duty to label (+)
Texts on topics of public interest
If someone uses AI to generate or modify texts that are aimed at the public and provide information on socially relevant topics, this must generally be disclosed. The aim is to protect trust in public debates and sources of information.
This refers to texts that deal with topics such as politics, society or culture and influence public opinion – particularly when they are disseminated via widely accessible channels such as social media, news outlets or websites and are aimed at a broad or easily accessible audience.
A text is subject to the disclosure requirement if:
- it has been generated or modified, in whole or in part, by AI, and
- the content is socially, politically, economically or culturally relevant, and
- the text is directed at the general public, e.g. via websites, social media, blogs or newspapers.
This does not apply to purely private or internal communications, such as emails to a small group or chat messages within a team.
Examples:
- A social media post about an election that was created entirely by ChatGPT
- A blog post on energy policy automatically generated by an AI platform
- An online comment on the Middle East conflict written and posted by a bot
- An AI-written article in an online magazine about the economic situation in Europe
- An automatically generated newsletter from an association on social debates
2. Who must fulfil the obligations?
Who is required to fulfil which transparency or labelling obligations depends largely on the role the respective company plays in relation to the AI system. The obligations under Article 50 of the AI Regulation apply to both providers and operators of the relevant AI systems:
- Providers: i.e. companies that develop AI systems or commission their development and then place them on the EU market under their own name or brand, or put them into operation under their own name or brand.
- Operators: companies that use AI systems under their own responsibility, for example to optimise internal work processes, in customer communication or in marketing.
Most companies that use common third-party generative AI tools such as ChatGPT, MS Copilot, Midjourney, etc. within their organisations are operators. Paragraphs 3 and 4 of Article 50 apply to them – that is, in the case of deepfakes and AI-generated texts of public interest.
However, it may also be the case that your company is a provider of an AI system, e.g. if your company has an AI system developed by a third party and then uses it under its own name for internal process automation. In that case, your company is subject to transparency obligations in relation to AI systems for direct interaction or in the case of synthetic content.
Further information on the roles of the provider or operator and the other stakeholders under the AI Regulation can be found here.
3. How must labelling be carried out?
Where labelling obligations apply, the information must be provided as follows:
- Timing: at the latest upon first interaction with or perception of the content
- Form: clear, unambiguous and transparent – not buried somewhere in the body text, in small print or hidden elsewhere on the website
- Accessibility: the notice must meet accessibility requirements, e.g. sufficient contrast, readability for screen readers, etc.
Depending on the context, appropriate wording could look like this, for example:
- Text: “This text was created with the support of AI.”
- Video: visible notice before start or displayed on screen
- Audio: an audio notification at the start – depending on the length of the audio content, this may also be repeated at intervals
4. Are there any exceptions to the labelling requirement?
Yes, there are also cases where labelling is not required:
Interactive systems (providers)
In exceptional cases, labelling is not required if it is clearly recognisable to reasonably informed and attentive users that they are interacting with AI. The recognisability must be obvious and context-specific.
For chatbots, labelling is generally required.
Synthetic content
There is no labelling requirement if the AI is only used in a supporting role and the content has not been substantially altered.
Examples: The AI suggests stylistic adjustments to a text, corrects spelling mistakes, translates or formats text.
Deepfakes
Deepfakes do not need to be labelled if:
- they do not relate to real people, objects, places, institutions or events, but to fictional representations, or
- it is clear and immediately apparent that the content is not based on reality, or
- the content in question is clearly of an artistic, creative, satirical or fictional nature – but only to the extent that labelling would impair the enjoyment or presentation of the work.
Check question: “Can the average user be misled by the AI output?”
Texts on matters of public interest
The labelling requirement does not apply if the AI-generated or AI-edited texts:
- remain internal or are made available only to a closed group of people,
- have no connection to the public interest, or
- it can be demonstrated that an employee checks the content and assumes editorial responsibility.
5. Recommendation for clients
- Introduce a central AI directory for content and communication systems
- Define fixed roles – e.g. the AI lead within the content team
- Draw up clear internal guidelines for AI labelling, including standard wording for text, images, video and audio, technical implementation on websites, in tools, CMS etc., and to ensure accessibility
- Introduce a review and verification process; particularly for content of public interest, this should include the following steps:
- Document editorial responsibility
- Verification process prior to publication (e.g. manual approval for AI-generated articles)
- Documentation of the distinction: public vs. internal
- Document exceptional cases or verification logs for unlabelled content
Would you like to set up your processes to comply with the AI Regulation now? Get in touch with us – we offer support on an individual basis or through workshops.