← Blog October 10, 2026

Invisible watermarks in ChatGPT: what the AI Regulation now demands from your business

OpenAI will invisibly mark ChatGPT text in the EU. What do the transparency rules of the AI Regulation mean for your chatbot, AI images and blogs, and why does December 2 matter?

Leeg vel papier met een vaag onzichtbaar patroon onder een vergrootglas op een strak bureau

On October 5, 2026, OpenAI announced that in the coming weeks it will add an invisible watermark to text generated by ChatGPT and Codex in the EU. That sounds like a technical detail, but it is a visible consequence of something bigger: the transparency rules from Article 50 of the European AI Regulation. These rules also affect ordinary SMEs, for example through a chatbot on the website, AI product photos or a blog written partly with AI. In this article, we outline what is changing, what is not changing and what you can do practically.

What OpenAI is exactly announcing

OpenAI will add an invisible watermark to suitable text from ChatGPT and Codex generated in the EU. For the API, it's different: watermarking is off by default there, and it will not become a global standard at launch either. API customers, wherever they are in the world, can turn it on themselves for certain models.

Anthropic went a step further in August 2026. New Claude models will receive an invisible watermark in generated text, based on a method derived from SynthID-Text, and according to secondary sources, Anthropic applies this globally rather than only in the EU.

What such a watermark cannot do

OpenAI is remarkably honest about the limitations. The watermark does not tell you who used the system, how much a human contributed, who owns the text, or whether the content is correct. And the absence of a watermark does not prove that a human wrote the text. Editing weakens the signal significantly: when a quarter of the words in English passages of 400 tokens were replaced with synonyms, detection dropped from approximately 92% to 17%. Furthermore, initially only approved researchers and expert organizations will have access to the detector.

The lesson: never treat a watermark check as proof, one way or the other.

What Article 50 means for your daily AI use

Article 50 has been in effect since August 2, 2026. Important to know: machine-readable marking of AI output is primarily a task for tool providers, such as OpenAI and Anthropic. For businesses using these tools, the main obligations lie elsewhere.

Hands with a smartphone next to a tablet and a card with an AI label icon on a dark workbench

The chatbot on your website

People need to know they are talking to AI. This information obligation has been fully in effect since August 2, 2026. A chatbot that pretends to be an employee, or where it is unclear that AI is behind it, is therefore not a good idea.

AI images and video

If you use AI for product photos, banners or videos, look especially critically at deepfakes: realistic-looking material that mimics people, places or events. That must be recognizable as such.

Blogs and social posts

Here is an important exception. A notice is only required for text published to inform the public about matters of general interest. And even then, it is not required if the text has been reviewed by a human or has undergone editorial control, and a person or company bears editorial responsibility. A human final editor is therefore not only good for quality, but also for your compliance.

Why December 2, 2026 matters

Many entrepreneurs think the AI Regulation has been postponed. That is partially correct. With the Digital Omnibus, which came into effect on July 27, 2026, the deadline for high-risk systems was moved to December 2, 2027, and the deadline for Annex I to August 2, 2028. However, the transparency rules of Article 50 remained unchanged.

According to the European Commission, AI systems that were on the market before August 2, 2026 have until December 2, 2026 to comply with the marking and detection obligation. New systems receive no transition period. Output generated before August 2, 2026 does not need to be retroactively labeled.

In the Netherlands, enforcement is somewhat behind. The draft Implementation Act for the AI Regulation came out of consultation on June 1, 2026, and the government accepts that the law will enter into force later than the AI Regulation requires. Opinions differ on who exactly will oversee it. The Dutch Data Protection Authority (AP) has already spoken out: organizations falling under these obligations should, according to the AP, study the Code of Practice and sign it in whole or in part.

Getting practical

On June 10, 2026, the Commission published the final Code of Practice, followed by guidelines on July 20. The Code is voluntary, but the transparency requirements of Article 50 are legal obligations. A few concrete steps:

Let your chatbot introduce itself. Make sure it is immediately clear to visitors that they are communicating with AI.
Appoint a human editor. Always have AI text reviewed by someone who takes responsibility for it.
Consider the EU icons. The Code of Practice contains free icons with 'AI', 'AI generated' or 'AI modified'. According to the Code, labels must be immediately recognizable without the user having to click anywhere, and noticeable upon first encounter.
Don't blindly trust watermark checks. Use them at best as one signal among others.
Inventory your AI use. Know which tools you deploy, what for, and who checks the output.

This is general information; if you have doubts about your specific situation, please seek legal advice.

At Silvar, we build with AI every day, from chatbots and document processing to AI videos and animations. We incorporate transparency from the design: a chatbot that honestly says what it is, and content that always has a human look at it. Would you like to discuss how to responsibly use AI in your own processes? We'd be happy to think it through with you.