Anthropic published a detailed blog post Friday explaining how its new text watermarking system will function across Claude models. The move responds to the EU AI Act's Transparency Code, which took effect August 2 and requires AI providers serving the European market to mark AI-generated content so other systems can identify it.

The company confirmed it will use a version of the SynthID-Text approach published by Google DeepMind in a 2024 Nature paper. The method alters the source of randomness when the model chooses between equally probable next words — such as picking "overcast" versus "grey" to describe weather — creating a statistical pattern detectable with a cryptographic key but invisible to human readers.

What's new

Anthropic's disclosure adds several concrete details not present in its earlier support page update. Watermarking will be applied at the model level, meaning it appears regardless of which Claude product or surface generates the text — including the Claude platform API, Claude Code, Claude Cowork, and Claude Tag. The watermark travels with copied and pasted text and may persist through some editing.

The company plans to release a watermark detection API, though no timeline was specified. Anthropic also emphasized that watermarking carries no identifying information and cannot be traced to a specific person, organization, or chat session. Other major model developers including Google, Meta, Microsoft, OpenAI, Black Forest Labs, and Synthesia have signed the same Code of Practice and will implement their own watermarks.

How it works and limitations

The watermark only activates on "low-stakes choices" where multiple word options are semantically equivalent. In factual passages where only one word is correct — such as "Isaac Newton's most famous work was called Principia Mathematica" — the watermark has nothing to act on. Similarly, when Claude only proofreads or lightly edits human-written text, nearly all words remain the human author's, leaving little for the watermark to attach to. Detection confidence increases with passage length.

Code generation receives lighter watermarking because working code requires specific syntax and logic, leaving fewer arbitrary choices. The watermark may still appear in comments and other areas where terminology choices are flexible, but Anthropic says the effect on actual code produced is negligible.

On the question of removal: light editing probably won't remove the watermark completely, while a complete rewrite replacing every word will — though Anthropic notes that in the latter case, "it's arguable whether the text can any longer be described as AI-generated."

Why it matters

The EU AI Act's Transparency Code represents the first major regulatory mandate for AI-generated content identification at scale. Anthropic's model-level approach — watermarking all outputs regardless of product surface — is more comprehensive than platform-level labeling alone. For developers and enterprises using Claude API, the watermark's persistence through copy-paste and partial editing means AI-generated content flowing into documentation, codebases, and publications will carry detectable signals.

This has implications for organizations with policies around AI disclosure, academic integrity, and content provenance. It also creates a practical test case for whether statistical watermarking can survive real-world workflows where text moves across platforms, gets edited, and is mixed with human writing.

Our take

Anthropic's transparency about limitations — especially on short texts, factual passages, and light editing — is notable. Most vendors would emphasize capability; Anthropic led with where the system doesn't work well. That honesty matters because regulators and platform operators will need to understand false-negative rates before building enforcement on top of these signals.

Sources