Your chatbot must introduce itself as an AI. Your avatar video needs a visible notice. And your automatically generated newsletter can no longer pretend that a human wrote it. Since the 2 August 2026 Do the transparency and AI labelling obligations of the EU AI Act apply to almost every organisation that uses AI? Sounds like bureaucracy? But above all, it's a question of orientation. And that is exactly what you get here – step by step.
What happened?
Article 50 of the AI Regulation (EU) 2024/1689 has been applicable since 2 August 2026. The idea behind it is simple and correct: People should be able to tell when they are dealing with AI – when they chat with a chatbot, watch a video or read a text. The rules apply directly in all EU member states, including Germany. A separate German law is not required for this.
Important to know: The obligations do not only apply to the providers of AI systems. Anyone who also uses AI appoints – as a business, public authority, association or sole trader – bears its own responsibilities. The European Commission published guidelines on this in July 2026, which clarify many practical questions.
Does that even concern me?
The shorts Answer: Very probably yes. Ask yourself three questions:
• Do you use a Chatbot or voicebot in customer service?
• Are you publishing Content that is AI were created – texts, images, videos, audio?
• Do you work with Avatars or AI voices, for example for training or marketing videos?
If you nodded even once, the labelling requirement applies to you as well. But – and that is the good news – it is quite manageable once you have understood the logic.
What you need to label:
• AI in direct contact: Chatbots, voicebots and virtual assistants must identify themselves – clearly and at the very latest at the beginning of the interaction.
• Deepfakes: Deceptively realistic AI images, voices and videos must be disclosed – with a visible overlay on videos, not just in the description.
• Unchecked AI texts on topics of public interest: Anyone who publishes AI-generated texts without a human having checked them and assumed responsibility for them must label them as such.
• Emotion recognition and biometric categorisation: Those affected must be informed. Emotion recognition is even banned in the workplace and in educational institutions.
What you do NOT need to label:
Here, most people breathe a sigh of relief – because much remains permitted without you having to do anything:
• Internal use: AI for drafts, research, summaries or translations does not trigger a labelling requirement. The decisive factor is publication.
• Editorially reviewed content: If a human reviews the AI draft, revises it and takes responsibility, the labelling requirement for the text no longer applies. By the way, that is exactly how we work – and that is exactly how you will learn it in our courses.
• Obvious AI: A clearly artificial robotic voice does not need any additional indication.
• Art and satire: A subtle, appropriate disclosure that does not disrupt the work is sufficient here.
How to correctly implement the AI labelling obligation – in 3 steps
Step 1: Get an overview. Where are you using AI – in customer contact, in marketing, in internal communications? Without this audit, you are operating in the dark.
Step 2: Decide for each piece of content whether it needs to be labelled. To do this, we have developed a decision tree with five simple yes/no questions – from „Is the content published at all?“ to „Has a human checked it?“. You will have your answer in under a minute.
Step 3: Use pre-written phrasing. There is no legally prescribed wording – but there are proven templates. For example: „Hello! I am the AI assistant from...“ for the chatbot, or „This video was created with AI – avatar and voice are artificially generated“ for the avatar video.
The only thing that matters is that the label belongs where the content is first perceived – not in the imprint.
And if I ignore it?
Breaches can be punished with fines of up to 15 million euros or 3 per cent of total annual global turnover – milder criteria apply to small and medium-sized enterprises. But honestly: the greater risk is the trust of your customers. Anyone caught concealing AI stands to lose more than just fine proceedings. Conversely: those who deal with AI transparently demonstrate competence – and build trust.
Whitepaper
Get the free white paper: „AI must be identifiable“
13 pages, free of legalese – created in collaboration with Marcel Helmcke, data protection specialist and lecturer at the Transmedial Akademie.
Inside you will find:
• the decision tree „Do I need to label?“ – in 5 questions to the answer,
• ready-to-use phrasing – for chatbots, telephone, images, video, audio, and text,
• 9 practical cases with step-by-step instructions – from website chatbot to public authority,
• the 10-point checklist for your business.
By the way: Labelling is only half the battle
Since February 2025, the AI Regulation requires something else: AI competence (Art. 4). Organisations must ensure that their employees are able to use AI competently. This is not a tedious duty, but a genuine opportunity – because anyone who understands how AI works not only labels properly, but also works better with it. That is precisely what our further training for is AI Prompter (Chamber of Commerce and Industry) – practical, up to 100 % eligible for funding via an education voucher, no prior knowledge required.
FAQ
Do I need to label every email I write with AI?
No. Internal and individual communication is not covered. The obligation applies to published content – and in the case of texts, only if no human has editorially reviewed and taken responsibility for them.
Does the obligation also apply if I only use ready-made tools like ChatGPT or HeyGen?
Yes. As a user, you are responsible for disclosing deepfakes and unverified AI-generated texts – regardless of who developed the tool.
Is a general notice in the imprint („We use AI“) sufficient?
No. The labelling must be clear, unambiguous and placed where the content is first perceived – on the image, at the beginning of the video, in the first chat message.
What applies to AI systems that were already in use before August 2026?
The machine-readable labelling obligations apply to them no later than 2 December 2026. The remaining obligations already apply.
This post was created with AI assistance and was editorially reviewed and is the responsibility of Phuong Tran, Transmedial Akademie. This is how we put into practice ourselves what we explain here. This post is for general information and does not constitute legal advice.
German 
English