Invisible Watermarks: How AI Texts Can Be Identified

With the Artificial Intelligence Regulation (the “AI Act”), the European Union has, for the first time, created a comprehensive legal framework for the use of artificial intelligence.

Since 2 August 2026, transparency and labelling obligations have applied to AI-generated content. Among other things, they require providers of generative AI systems to technically label AI-generated or AI-manipulated content in such a way that its artificial origin can be identified in machine-readable form. In addition, certain AI-generated or AI-manipulated content published by companies is subject to labelling obligations vis-à-vis users.

This raised the practical question: How is such technical labelling supposed to work for AI-generated texts?

Anthropic announced that it would equip future models with watermarking for texts.

What is “watermarking“? 

When one thinks of watermarks, one initially thinks of visible indicators, such as logos, patterns, or lettering, that identify the origin or author of a document.

In AI-generated texts, however, watermarking operates through a statistical pattern in the choice of words that is not discernible to the reader.

How does “watermarking” work?

Generative AI produces texts by having the underlying model decide, step by step, which word is most likely to come next. There are often several words that fit equally well.

Example: “The film was quite…“ 

Here, “exciting“ or “captivating“ could follow. Both words fit well both linguistically and in terms of meaning. Which variant is chosen makes hardly any difference to the meaning of the sentence.

These “low-stakes decisions“ are used in watermarking: If several equally suitable words are available, the random selection among them is controlled so that a particular statistical pattern emerges over the course of a longer text.

In doing so, the model does not alter the content of the text or add any hidden characters or additional words.

The watermark is therefore not discernible to the reader. Anyone who has the relevant key, however, can check whether the text exhibits the expected statistical pattern.

The limitations of “watermarking“

A watermark does not prove that a text was created entirely by AI. Rather, it may indicate how likely it is that a particular AI system was involved in creating or editing the text.

The following applies: The fewer textual decisions the AI model makes itself, the fewer opportunities there are for “watermarking.”

If AI merely corrects a text written by a human, most of the words still originate with the human. The opportunities for incorporating “watermarking” are therefore more limited than with a text generated entirely by the AI itself.

Even with unambiguous answers, there is no room for choice of words.

Example: “2 + 2 = …“

There is no equivalent alternative to “4” here. Watermarking cannot be applied in this case. 

What does this mean for companies?

The question of whether and how AI-generated content must be labelled is an important compliance issue.

Anyone using AI for marketing, product descriptions, customer communications, or other processes within the company should be sure to clarify the following three questions:

1. AI detection:

Are contents created or edited using AI, and can this be technically verified?

2. Transparency obligation:

Must the content be labelled for users as AI-generated or AI-manipulated?

Technical labelling, such as Anthropic’s watermarking, is not the same as a visible disclosure to the user.

3. Lawfulness of the content:

Is the company permitted to use the content at all, for example with regard to copyright, personality rights, or competition law, and who is responsible for its use?

Technical watermarks can help with the first question. However, they do not automatically answer the second question, and certainly not the third. Companies should therefore not only know whether they are permitted to use AI, but also which transparency obligations and other legal obligations are associated with the specific use.