Invisible watermarks won't prove text authenticity, they just comply with the AI Act

A European whip unifies infrastructure

Anthropic has announced it will start adding invisible watermarks to text generated by Claude. The move is a response to obligations under the European Union AI Act, which requires machine readable labeling for synthetic content. Regulation is thus forcing competing companies to reach for the same tools. It is no coincidence that Anthropic is using a version of the SynthID approach, an open technology from Google DeepMind. Google itself has had this deployed in Gemini since 2024.

Statistics are not bulletproof

The method does not work like invisible ink, but as a statistical tweak to the probability of low stakes words. For example, whether the model chooses overcast or grey. This imprints a specific pattern into the text without affecting the API price or the practical quality of the responses. However, such a watermark only survives in longer blocks of unchanged text. It inherently cannot be robust.

Every edit destroys the evidence

The moment a user deletes a copied paragraph, rewrites a few words, or even runs it through another model for paraphrasing, the detection signal reliably breaks down. Watermarking will help automated scanners catch massive, raw slop on the web, but it cannot stop deliberate forgery. If a document is meant to influence a court, an election, or corporate decisions, the person generating it will always take the five minutes to overwrite the invisible pattern.

A shared key will show willingness for oversight

The main signal in the future will not be the deployment of SynthID across other models, but whether the market agrees on a unified detection API. If OpenAI, Anthropic, and Meta each build their own detector without central verification, users will have to run texts through several different checks. The success of watermarks paradoxically depends on a shared reading key, not on who invents a more complex number generator.

Lilith's verdict

The bureaucratic checkbox is ticked. Now someone just needs to explain to regulators that an invisible text watermark can be broken by an intern with a backspace key and a thesaurus.