Claude Text Watermarks: How They Work and Why They Went Global

A pattern in word choice, not a visible tag

Anthropic's watermark is built on SynthID-Text, a technique Google DeepMind published in Nature in 2024 that traces back to a 2022 proposal by Scott Aaronson. Rather than inserting visible characters or altering the meaning of a response, it changes the source of randomness the model draws on when it settles which word to pick next, using an encoded key and the preceding few words. The choice at each step still looks random, but a detector holding the key can recover the pattern. [1]

Anthropic says the mark travels with the text when it is copied and pasted elsewhere and can persist through some editing, and that it applies across Claude's products: the API, the Claude apps, Claude Code, Claude Cowork and Claude Tag. Files get a separate mechanism, the C2PA open standard, rather than SynthID-Text. [2]

Anthropic is explicit that detecting the mark is not proof: it can only indicate that Claude was likely involved with a piece of text at some point, and it cannot distinguish text Claude wrote from text Claude heavily edited. Detection needs enough length and enough creative latitude in word choice to work at all, so factual passages and code carry little to no watermark signal. [1]

New models first, worldwide from the start

The watermark applies automatically to Claude models released from August 2, 2026 onward. For models that predate that cutoff, Anthropic says it is working to add watermarking over the coming months rather than all at once. [1]

Anthropic applied the watermark globally rather than confining it to the European Union, stating that it does not yet have a durable way to scope the mark by region, while saying it would keep evaluating regional approaches. [1][3]

Atlas interpretation: That worldwide-by-default choice is the more consequential part of the announcement, since a rule written for one jurisdiction ends up governing every user of the product regardless of where they are. Building region-specific behavior into a model's output pipeline is evidently harder than switching a feature on globally, at least on Anthropic's own account of why it did not scope the rollout. [1]

The EU rule arrived first

The EU AI Act's transparency obligations took effect August 2, 2026, requiring that generated content carry a machine-readable mark identifying it as artificial. Anthropic signed the EU's Code of Practice on Transparency of AI-Generated Content, one of roughly 190 signatories that also include Black Forest Labs, Google, Meta, Microsoft, OpenAI and Synthesia. Nine days after the rule took effect, on August 11, Anthropic announced the watermark by updating its help-center page on how Claude marks AI-generated content; the longer explainer on its news site followed on August 14. [1][2][3]

Atlas interpretation: TechCrunch's reporting describes the watermark as a direct response to that EU deadline, which lines up with the timing: the code of practice was signed in July 2026 and the enforcement date landed August 2, with Anthropic's announcement following nine days later. Anthropic's own announcement does not dispute that reading, though it presents the global scope as a practical limitation rather than a policy choice to extend EU-style disclosure everywhere. [2][1]