AI Law & Compliance
AI labeling obligation: What applies from 2 December
On 2 December, a grace period under the EU AI Act ends. What has applied since August, who has to do what, and how Kasimir takes work off your hands.
Monday morning in the marketing team. A colleague creates an image for the new LinkedIn post with AI, uploads it, and the platform automatically adds “Made with AI” underneath. How does LinkedIn know? And who should have labeled it in the first place: her, the company, or the tool that created the image?
These are exactly the questions Article 50 of the EU AI Act answers. And since 2 December is circulating everywhere as the deadline, it is worth taking a closer look.
- Already in force since August
The labeling obligations in Article 50 have applied since 2 August 2026.
- 2 December is a grace period
By then, providers of existing AI systems must mark their output in a machine-readable way.
- You have duties of your own
Disclose deepfakes and label certain published texts.
Why the EU wants this
AI can now generate text, images, voices and videos that are hard to tell apart from the real thing. That is useful when you draft an offer or illustrate a presentation. It becomes a problem when fake recordings of people appear or automatically generated news shapes opinion. The AI Act therefore relies on transparency: anyone who encounters a piece of content should be able to tell that an AI created it. For people through visible notices, for machines through a marking that platforms and checking tools can read.
What really happens on 2 December
In July 2026, the AI Act was amended by the so-called Digital Omnibus. It contains a transitional rule: providers of AI systems that generate synthetic audio, image, video or text content and were placed on the market before 2 August 2026 must meet the machine-readable marking obligation by 2 December 2026. The other obligations in Article 50 were not postponed.
Who has to do what? Providers and deployers
The AI Act distinguishes two roles. Providers develop an AI system and place it on the market. Deployers use it under their own responsibility, which describes almost every company that uses AI in its daily work.
What your AI provider does:
It makes sure people know when they are chatting with an AI.
It marks generated text, images, audio and video as AI-generated in a machine-readable way.
What you do:
You disclose deepfakes: AI images, videos or audio that look authentic and resemble persons, objects, places or events that exist or could plausibly exist. For evidently artistic, satirical or fictional works, a notice that does not spoil the work is enough.
You label AI-generated texts that are meant to inform the public on matters of public interest. This does not apply if a person has reviewed the text and someone holds editorial responsibility for it.
If you build your own chatbot for customers or your own AI tool, you may also be a provider. When in doubt, a short legal review is worth it.
Four everyday cases
An AI illustration for your website. If it is clearly drawn or stylised, it needs no notice. If it looks like a real photo of people, products or places, it is a deepfake according to the European Commission’s guidelines, even if the person shown does not exist. Then it needs a notice. The provider’s marking is embedded in the file. Do not remove it, for example through image editing or export settings.
A press release drafted by AI. No notice is needed if someone on your team reads it, approves it and takes responsibility for it. That review is exactly the exception the law provides.
A summary for an internal meeting. No notice is needed. Internal documents do not inform the public.
A chatbot in customer service. Yes. Anyone writing there must be able to tell that an AI is answering, unless it is obvious anyway.
How Kasimir handles it
As the provider of Kasimir, the marking obligation is ours, not yours. This is how we meet it:
Images that Kasimir generates carry a machine-readable marking following the IPTC standard for AI-generated media. When exporting, you can also add a visible watermark, and your company can make it mandatory for everyone.
Documents: Word and PDF files from workflows with an AI step, meeting summaries and texts downloaded as Word additionally carry a digitally signed marking in their metadata.
Chat answers are recognised by our checker: at kasimir.ai/en/ki-pruefen anyone can check for free whether a text or file came from Kasimir. We do not store what is checked there.
Source references make it easy to review texts before publishing. That human review is also the exception from your notice obligation for published texts.
Operated in Germany: Kasimir is operated in a German data center. Frontier models run GDPR-compliant in the EU, and your content is not used for training.
Your checklist for the coming weeks
List the AI tools you use and ask each provider how it will implement the marking by 2 December.
Check chatbots on your website and in customer service: is it clear that an AI is answering?
Decide who reads and approves AI-generated texts before they are published.
Visibly label realistic AI images and videos of people before they appear in advertising or on social media.
Make sure image editing and exports keep the metadata intact.
Write the rules down in a short AI policy and train everyone who publishes content.
Three questions we hear often
Do I have to label every email I wrote with AI?
No. According to the European Commission’s guidelines, emails, offers and internal documents are not considered published. The notice obligation only covers texts meant to inform a broad public on matters of public interest.
What about AI translations?
The European Commission’s guidelines, which are not legally binding, treat translations, spelling corrections and transcripts as “standard editing”. The provider’s marking obligation does not apply to them. Summaries and texts that change style, structure or meaning, on the other hand, must be marked.
What happens if a label is missing?
The regulation provides for fines of up to 15 million euros or 3 percent of worldwide annual turnover, whichever is higher. For small and medium-sized companies, the lower of the two amounts applies.
This article provides an overview and is not legal advice.
Regulation (EU) 2024/1689 (AI Act), Articles 50 and 99
European Commission: Code of Practice on marking and labelling of AI-generated content
European Commission: Guidelines on the transparency obligations under Article 50 (20 July 2026)
GDPR-compliant AI from a real German data center
Kasimir runs on its own infrastructure in Germany — no detour via US providers, no CLOUD Act reach.
