Will Claude’s invisible watermark change your brand’s tone of voice?

At first glance, Anthropic’s new text watermark sounds like a compliance feature rather than a creative one. Introduced in connection with the EU AI Act’s transparency requirements, it is designed to leave an imperceptible, model-level signal in text produced by Claude. Anthropic has confirmed the watermark, although it has not disclosed exactly how its implementation works.

But if your organisation cares about your AI generated content maintaining brand tone of voice, there is an awkward consequence: the watermark will change it.

Statistical watermarking for large language models works by adjusting the probabilities the model uses when choosing its next word (token), often by softly favouring a changing subset of the available options. This means that when a model-level watermark creates a detectable pattern in the text, it must influence at least some of the model’s choices. If it never changed a choice, there would be no textual signal to detect.

And word choice is not separate from tone of voice; it is one of its essential ingredients, alongside sentence structure, rhythm, formality, humour and point of view. Change the probability of one word over another, and the model will select a phrase that is less characteristic of the brand’s preferred language.

Anthropic have described these changes as being “imperceptible”. That distinction matters because “imperceptible” likely means no noticeable loss of quality. A paragraph can remain perfectly fluent while becoming less faithful to a brand tone of voice.

Brand teams are not merely asking whether copy is good; they are asking whether it sounds recognisably like us.

The watermark also needs to be understood for what it can and cannot establish. Anthropic says detection indicates that text has been processed by Claude, not that Claude was its sole author, and the signal may remain after tasks such as proofreading or translation. Conversely, the absence of detection does not prove that AI was not involved. This makes watermarking a useful transparency mechanism, but a poor substitute for provenance, editorial judgement or sensible governance.

Time Under Tension is a Microsoft partner and OpenAI Services Partner, and we also do substantial work with Anthropic models. This is not an argument against Claude, which is often an excellent choice. It is an argument for treating model selection as a brand decision as well as a technical and commercial one.

Organisations should now test leading models against a fixed set of real brand tasks, including difficult examples where vocabulary, rhythm and point of view matter. Build a brand voice test set, compare outputs systematically, and retain human editorial review for important material. Training should help people recognise voice drift, platform selection should account for it, governance should document it, and AI workflows should include a deliberate editorial checkpoint.

Watermarking may improve transparency, but it is not creatively neutral. If the mechanism changes textual choices, it will sometimes change the brand voice those choices create, even when almost nobody can see it happening.