Anthropic embeds invisible watermarks in Claude text outputs worldwide

Anthropic has added invisible watermarks to text generated by its Claude models, and the marking applies to output worldwide rather than only in the EU. The trigger is the European Union's AI Act: Article 50's transparency provisions became applicable on August 2, 2026, requiring providers of generative AI systems to mark outputs in a machine-readable way so content can be detected as artificially generated. Anthropic's documentation confirms the marking ships with supported models in every region where Claude is offered.

The architectural reason for the global rollout is that the watermark operates at the model level, not the account or region level. Once embedded in the model's sampling behavior, it applies uniformly. Models launched after August 2 carry the mark from day one; systems already on the market have until December 2, 2026 under a grandfathering window.

The technical approach differs by medium. Images receive C2PA provenance metadata, which is fragile — a format conversion or screenshot strips it. Text watermarks work differently: the mark is embedded in word-choice patterns within the text itself, so it survives copy-paste and may persist through light editing. Detection is a statistical confidence score across a passage, comparing word choices against chance. Short passages don't carry enough signal for reliable detection, and substantial rewriting degrades the signal significantly.

The counterintuitive consequence: documents most likely to be flagged are the ones that received the least human editing.

Current detection is binary — it reports whether a mark is present, not what proportion of a document was AI-generated. Anthropic has said the detection mechanics are still being documented, so the tooling should be treated as unfinished. Published research is moving past it: sliding-window scoring methods can locate contiguous marked blocks that whole-document scoring misses, and work exists on estimating the share of a mixed document that came from a model, using legal briefs as a worked example.

Anthropic is not alone in signing the European Commission's voluntary Code of Practice on Transparency of AI-Generated Content. OpenAI, Google, Meta, Microsoft, and Mistral are among the roughly 190 signatories. But Anthropic has published more detail on its text watermarking approach than its competitors; OpenAI in particular has been historically reluctant about text watermarking over accuracy concerns.

The watermark carries no prompt content or client information. It signals only that a model processed the text, not that it authored it. And deliberately stripping a mark is not merely a technical workaround: Article 50 prohibits removing or altering AI markings, so doing so converts a passive fact into an affirmative act with legal consequences.

The open question is whether today's binary detection and its accuracy limits represent a stable endpoint or a temporary one.