An invisible watermark in text: how companies meet their AI Act duties

Marking AI-generated content sounds like a question of honesty toward the reader. In practice it is a technical problem, and the industry handles it in different ways. heise online describes how Anthropic is extending text watermarking to more Claude models, with the EU artificial intelligence regulation as the backdrop.

According to heise, from 30 September 2026 the invisible watermark will also go on the output of Claude Fable 5, Claude Sonnet 5 and Claude Opus 4.8. Claude Fable 5.1, Claude Mythos 5.1, Claude Opus 5.5 and Claude Opus 5 already carry it. The last of these was marked in a post-launch update. Older models are to be covered by 2 December. The information comes from an email Anthropic sent to customers on 24 September. One detail matters: the mark is applied at the model level, and the user cannot switch it off.

How it works

The method does not rely on hidden characters or metadata, which are easy to remove. Anthropic uses its own variant of the SynthID-Text technique developed by Google DeepMind. The mechanism influences the choice of one of several equally likely next tokens. Over a longer text a statistical pattern emerges. The right key can detect it, provided the passage being analysed is long enough. The company says this does not degrade text quality.

The limitations are more serious than the simplicity of the idea suggests, and Anthropic itself admits as much. A match indicates only that Claude probably took part in producing the text. The model can rewrite someone else's passage, add its own sentences or translate another person's text, and the mark disappears. So this is not an evidentiary tool but a signal of probability.

Where the obligation comes from

The reason is regulatory. The EU AI Act imposes transparency duties on the builders of generative systems. Since 2 August they have had to ensure that content generated by AI, including text, is marked in a machine-readable way and recognisable as generated. Anthropic applies its solution globally, across all Claude platforms, even though the initiative came from Europe.

There is also an institutional context. The European Commission runs an AI Office, more than 125 staff across six units, which handles the implementation of the AI Act, including the rules for general-purpose models. Among other things it can carry out evaluations of such models, demand information and impose sanctions. The marking obligation is therefore only one of many requirements that model providers have to factor into their production process.

Does it work at all

Criticism comes from two directions. On one side, practitioners point out that text watermarks are vulnerable to ordinary editing and easy to remove by translation or summarisation. On the other, the opposite question arises: what if a mark is detected in a text that a person wrote themselves, using a model only for proofreading? Then the meaning of "marked as AI" blurs and a false accusation becomes easy. The safest reading of this mechanism is as a tool for statistical monitoring of content at the scale of the internet, not for settling the authorship of a single paragraph.