AI watermark

How a watermark gets into the text

When generating, a language model chooses at each position from several possible next words. A watermarking procedure shifts this choice minimally: certain words are used slightly more often, others slightly less often, than would be the case without the marker — following a pattern that can only be reproduced with the associated key.

Taken individually, each of these deviations is meaningless and indistinguishable from chance. Only across a hundred or several hundred words does a pattern emerge that becomes statistically significant. This is why text watermarks work poorly or not at all on short texts.

The best-known documented procedure is SynthID Text from Google DeepMind, published in the journal Nature in 2024 and available as an open-source tool.

Can an AI watermark be recognised?

As an outsider, generally not. A watermark is only of use to whoever holds the key to read it — normally the provider that set it. To date there is no common standard and no generally accessible verification tool with which arbitrary texts could be checked for arbitrary watermarks.

It is important to distinguish this from the so-called AI detectors offered online. These do not read out a watermark; they estimate from stylistic features whether a text originates from a machine. Their error rate is high, and they regularly flag texts written by humans. A watermark verification and a detector estimate are two fundamentally different things.

Can a watermark be removed?

In theory yes, in practice with some effort. Since the signal sits in the word distribution, every change weakens it: rewriting, shortening, reordering sentences, adding your own passages. Research on the reliability of such procedures shows that they are robust against light paraphrasing, and considerably less so against thorough revision.

From this follows a rather unspectacular insight: anyone who reworks a text far enough for the marker to disappear has invested roughly the effort that proper editorial work costs anyway — and usually with a poorer result, if obscuring the origin was the sole purpose.

What the EU AI Act requires

The relevant provision is Article 50 of the EU AI Act, whose transparency obligations apply from 2 August 2026. It contains two obligations that are frequently confused:

• Providers of AI systems that generate text, images, audio or video must mark the outputs in a machine-readable format, so that it is recognisable that they were artificially generated.
• Deployers who publish AI-generated text in order to inform the public about matters of public interest must disclose that the content was artificially generated — unless a human has reviewed it editorially and bears responsibility for it.

An obligation to provide every text on your own website with a watermark does not follow from this. The labelling obligation applies to the model providers, not to the website operator.

Relevance for SEO and GEO

At present, none that can be evidenced. Google states that AI-generated texts are not, as such, a ranking disadvantage; what matters is quality. Whether search systems read origin signals at all is not publicly confirmed — and technically it would require them to hold the key of the respective provider, or for a common standard to emerge.

In practice this means: a content strategy built on concealing origin is unprotected against any future change of rules. Anyone who takes editorial responsibility for texts and makes that visible is in a better position both legally and in terms of content.