TL;DR
- Germany's search term for this is KI-Kennzeichnungspflicht, and it is not one duty but four, all in Article 50 of the EU AI Act (Regulation (EU) 2024/1689): chatbot disclosure, machine-readable marking of AI output, notice for emotion recognition, and a visible label on deepfakes and AI-written text on matters of public interest.
- All four apply since 2 August 2026. The Digital Omnibus, Regulation (EU) 2026/1744, did not move that date.
- One exception: generative systems already on the market before 2 August 2026 must deliver the machine-readable marking under Article 50(2) only from 2 December 2026. That is the Omnibus transition. New systems are covered from day one.
- Fines sit in the second tier: up to EUR 15 million or 3% of worldwide annual turnover (Article 99(4)). There is no grace period for the chatbot notice or the deepfake label.
- For most companies the work is small: a visible notice on the chatbot, a label on realistic AI images and video, and, if you ship generative software, a technical mark on its output.
Which labelling duty applies to AI content from 2026?
People asking this usually mean two different things: telling someone they are talking to an AI, and the label on an AI-generated picture. Both live in Article 50, alongside two duties that bite less often.
Role decides everything. A provider develops the system or puts it on the market under its own name. A deployer uses it professionally. Most companies are both: deployer of the Intercom or Tidio bot on their website, provider of the image generator they built into their product. Get the role wrong and you satisfy the wrong duty. The provider vs deployer guide works through it.
The two dates: 2 August 2026 and 2 December 2026
This is where most of the confusion sits, and the answer is short.
Article 50 applies since 2 August 2026. The Digital Omnibus, in force since 27 July 2026, deferred the high-risk deadlines (Annex III to 2 December 2027, Annex I to 2 August 2028). It left the transparency duties alone. A chatbot running without a notice in autumn 2026 is non-compliant now, not next year.
2 December 2026 concerns only Article 50(2), and only legacy systems. The Omnibus carries a transitional provision: generative AI systems already placed on the market before 2 August 2026 must deliver machine-readable marking of their output from 2 December 2026. Anything new since 2 August 2026 is covered immediately.
Real-world example: a SaaS vendor has offered a feature since 2025 that drafts product descriptions with a language model. Marking that text under Article 50(2) has until 2 December 2026. The chatbot the same vendor adds to the app in September 2026 needs the Article 50(1) notice from day one. Two duties, two clocks, one product.
Four months is not long for a technical mark that has to survive screenshots and re-encoding. If you use the transition, document that you are on schedule. Before relying on the December date for a specific system, read the Official Journal text of Regulation (EU) 2026/1744 itself.
Do I have to label my AI chatbot?
Yes, in almost every case. Article 50(1) requires that people are told they are interacting with an AI unless that is "obvious from the circumstances". The exception is narrow. A fluent support bot with a first name does not qualify.
A notice that holds up:
- Timing: at first interaction, not after the third message.
- Place: on the chat widget itself, not in the terms or the privacy policy.
- Language: every language the bot serves, and accessible, so not only as an image or a colour.
- Wording: not prescribed. "You are chatting with an AI assistant. Replies are AI-generated." is a solid baseline.
If you run a third-party bot you are the deployer, and you are responsible for the notice actually appearing on your site. The full answer with the edge cases is in Do I need to disclose my AI chatbot?. The free Article 50 disclosure generator produces the wording, no account needed.
How do I label AI-generated images?
For images, video and audio there are two layers, and they get mixed up constantly.
Layer 1: machine-readable, the provider's duty (Article 50(2)). The system that generates the content must mark its output technically as artificially generated: readable by detection tools and platforms, usually invisible to people. Neither the Regulation nor the Commission's guidelines of 20 July 2026 prescribe a technology. The Code of Practice on marking and labelling AI content, final since 10 June 2026 and voluntary, describes a layered approach, because any single mark can be stripped in normal distribution:
C2PA is the obvious reference point because Adobe, Microsoft, Google, OpenAI and camera makers support it, not because the law names it. How the layers fit together is on the content marking page.
Layer 2: visible, the deployer's duty (Article 50(4)). Whoever publishes a deepfake must disclose it in a way people can see. Deepfake under Article 3 is broad: AI-generated or manipulated image, audio or video that resembles real persons, places, objects or events and would falsely appear authentic to someone. That covers AI product photos in realistic scenes, cloned voices in marketing, and AI profile pictures in user-facing contexts. The label has to be clear, at the point of use, and effective for the medium: visual for image and video, audible for audio.
Real-world example: an online shop generates lifestyle photos of its furniture in realistic living rooms. Those images need a visible label ("AI-generated"), because they simulate a real scene. The abstract background pattern on the homepage, also generated, needs none: it resembles no real person, place or event.
For AI-generated text the visible label applies only where the text is published to inform the public on matters of public interest and has not passed human editorial review with a natural or legal person holding responsibility. An AI-assisted product description approved by a person is out of scope.
Is AI labelling mandatory on Instagram?
Two rulebooks that have nothing to do with each other:
- The legal duty in Article 50(4) applies regardless of platform. Publish a deepfake on Instagram, LinkedIn or your own site and you must label it. A platform label only substitutes for your own disclosure if it is clearly visible to the viewer.
- Platform rules live in each service's terms. They can be stricter than the law and reach content that is not a deepfake. They change nothing about Article 50.
Practically: label AI images and video at creation, with a visible label and with metadata intact, rather than relying on platform detection. Platforms routinely strip metadata on upload, and a missing label is then your problem, not theirs.
Is an AI policy mandatory?
Article 50 does not require a policy, it requires notices and marks. A written AI policy is still the usual way to evidence two other duties: AI literacy under Article 4, in force since 2 February 2025, and the record of who in your company applies which label and when. When an authority asks, "we have a policy, here it is, here are the training records" is the shorter conversation.
What non-compliance costs
Article 50 breaches sit in the second fine tier of the AI Act: up to EUR 15 million or 3% of worldwide annual turnover for the preceding financial year, whichever is higher (Article 99(4)). For SMEs and start-ups the lower of the two amounts applies. Enforcement is by the national market surveillance authority.
There is no transition for the chatbot notice or the deepfake label. The only transition is the one above, for machine-readable marking by legacy systems.
Checklist: the labelling duties in 60 minutes
- Inventory (10 minutes). Which AI talks to people (chatbots, voice agents, assistants)? Which generates content (text, image, audio, video)? Which did you build, which is third-party?
- Role per system (5 minutes). Provider or deployer. If in doubt, both.
- Chatbot notice (15 minutes). Notice on the widget, at first interaction, in every language. On WordPress the EU AI Act plugin does it locally with no account; the disclosure generator supplies the text.
- Visible labels (10 minutes). One rule for marketing and content: realistic AI images, video and voices are labelled before publication.
- Machine-readable marking (planning, 15 minutes). Only if you ship generative software yourself. Check the date: new since 2 August 2026, or legacy with the 2 December 2026 deadline. Pick the approach (content credentials, watermark, metadata) and document the timeline.
- Evidence (5 minutes). Screenshots of the notices, dates, owners. That is what you show an authority. Templates are in the Article 50 Disclosure Pack.
FAQ
Which labelling duty applies to AI content from 2026?
Since 2 August 2026, the four duties in Article 50: chatbot disclosure, machine-readable marking of AI output, notice for emotion recognition, and a visible label on deepfakes and AI text on matters of public interest. For machine-readable marking by systems on the market before 2 August 2026, the date is 2 December 2026.
Since when does the AI labelling duty apply in Germany?
Since 2 August 2026, directly from the AI Act. No German implementing law is needed; an EU regulation applies as it stands.
Do I have to label AI-written text on my website?
Visibly only if the text is meant to inform the public on matters of public interest and appears without human editorial responsibility. An approved blog post or product description is out of scope. If you generate the text with your own software, the machine-readable marking under Article 50(2) binds you as provider.
Is a line in the privacy policy enough?
No. The chatbot notice must appear on the bot at first interaction, the deepfake label on the content. A sentence in the privacy policy or the terms satisfies neither.
Does it apply to small companies and freelancers?
Yes. Article 50 has no size threshold. The difference is the fine: for SMEs the lower of the two amounts in Article 99 applies.
Which technology must I use for machine-readable marking?
None in particular. The Regulation, the guidelines and the Code of Practice are technology-neutral. C2PA Content Credentials are the common reference for image, video and audio, metadata the practical route for text, a watermark the robust second layer.
Do it now
- Generate the notice text (free, no account): Article 50 disclosure generator
- Chatbot notice on WordPress: EU AI Act plugin · install from WordPress.org
- Understand machine-readable marking: Content marking
- Check every duty: Free EU AI Act assessment
Sources
- Regulation (EU) 2024/1689 (AI Act), Official Journal, EUR-Lex
- Regulation (EU) 2026/1744 (Digital Omnibus), Official Journal, EUR-Lex
- European Commission: Guidelines on Article 50 transparency obligations (20 July 2026)
- European Commission: Code of Practice on transparency of AI-generated content, how to sign
- European Commission: AI Omnibus enters into force
This post is general information and not legal advice. For your specific case, consult a qualified professional.
