From 30 September 2026, Anthropic will apply an invisible text watermark to outputs from further Claude models. According to the announcement, the models affected are Claude Fable 5, Sonnet 5 and Opus 4.8. The company cites the transparency requirements of the EU AI Act, the European regulation on artificial intelligence, as the reason.
Technically, this is not a visible mark in the text, nor is it additional metadata that could simply be removed. Instead, a statistical pattern is created in the choice of words: during generation, the model favours certain phrasing variants in such a way that a signal can be derived from them computationally. For readers, the text remains inconspicuous.
At model level, not as an add-on feature
What matters in practice is where it is implemented. The watermark is embedded at model level and therefore applies across platforms, regardless of whether the output is produced via a web interface, an app or an API. Customers cannot disable the feature.
This distinguishes the approach from labels that are only added downstream in a product. Anyone integrating Claude into their own systems via an API therefore receives the signal automatically, without having to configure anything for it — and without the option of opting out.
Why this is connected to the AI Act
The EU AI Act requires, among other things, transparency for content generated or altered by AI systems. Providers are to make it machine-readably traceable that content has been synthetically produced. A watermark at model level is one way of not leaving this traceability to the individual use case, but shifting it into generation itself.
For companies that produce AI texts, this shifts part of the responsibility: the technical marking is supplied by the model provider. Whether and how content is labelled for users, by contrast, remains a decision within their own process.
What changes for website and shop projects
In the everyday life of many projects, AI-assisted texts are created in very different places: product descriptions in the shop, meta texts for search engines, drafts for blog posts, suggested replies in support, translations, summaries of longer documents. These texts often pass through several hands and systems before they are published.
This is exactly where a sober stocktake is worthwhile. Sensible questions are:
- Where in the editorial or shop system are AI-generated contents created, and who initiates them?
- Does the system record which model generated a text and how heavily it was subsequently edited?
- Is there a defined approval step before generated content becomes publicly visible?
- How is content labelled for the audience when labelling is required or desired?
In TYPO3 and comparable systems, such information can be kept as additional fields on content elements or records. This costs little to implement, but afterwards makes the difference between a robust statement and an estimate.
Assessing the limits of detection realistically
A watermark is a signal, not proof in the legal sense. Texts are shortened, rewritten, translated, mixed with passages of one's own. The more heavily an output is editorially revised, the less of the original statistical pattern remains. Conversely, a detected signal says nothing about whether content is factually correct.
Anyone wanting to build an internal control logic from this, for example to check supplier texts, should do so with restraint. It is more sensible to document origin and depth of editing cleanly within one's own process, rather than trying to reconstruct them later from the finished text.
The contractual and organisational side
The fact that the feature cannot be switched off also has a procurement dimension. Companies that embed AI outputs in products or pass them on to customers should know what properties the models used have and how this affects their own assurances. This applies to agency contracts just as much as to software that delivers generated content to third parties.
In practice this means: model selection and model changes should be documented. If a project relies on one model today and on another in a year, it should remain traceable which content was created under which conditions.
Assessment
The labelling of AI content is moving from the application layer into the model itself. For web and software projects, this means less scope for shaping the technical marking, but also less effort to produce it themselves. The decisive part remains organisational: anyone who now makes origin, depth of editing and approval of generated content traceable within their own system is considerably better prepared for future proof obligations than through retrospective reconstruction.