Anthropic announced on August 14, 2026 that future Claude models will generate text containing a watermark intended to indicate the likelihood that Claude was involved in producing the text. The change is being implemented alongside other major AI providers to comply with the EU AI Act and the EU Code of Practice on Transparency of AI-Generated Content signed in July 2026 (about 190 total signatories).
How the text watermark works — the basic idea
Large language models like Claude generate text token by token. In many positions multiple candidate tokens would be acceptable; the model’s selection among those candidates is often random or near-random. Watermarking leverages those low-stakes choices to leave a detectable pattern. The reader should not perceive any difference: the watermark alters the source of randomness used to pick among equivalent options, replacing an arbitrary random number generator with a function that depends on a secret key and the preceding few words.
From the user’s perspective the words remain plausible and the meaning is unchanged. From the perspective of someone holding the watermark key, the sequence of tokens can be analyzed to assign a probability that the text was generated by Claude using that key.
The method used
Claude’s text watermark is a variant of the SynthID-Text approach published by Google DeepMind in Nature in 2024, and traces back to a proposal by Scott Aaronson in 2022. All these methods share the same design principle: change the randomness source rather than inserting visible markers or hidden characters into the text.
Effect on outputs, speed, and cost
Anthropic reports no practical impact on output quality, creativity, or readability. Internal testing and prior evaluations of SynthID-Text (including a traffic experiment by Google DeepMind) found no statistically significant quality differences between watermarked and unwatermarked outputs. Because the watermark does not add extra tokens, it has negligible effect on latency and does not increase serving cost.
Privacy and identifiability
The watermark carries no identifying information about individual users, organizations, or chats. It only indicates whether Claude likely participated in producing or processing the content; it cannot be traced back to a specific person or conversation.
Limitations and edge cases
- The watermark can only estimate the likelihood that Claude was involved; it cannot definitively prove authorship or distinguish between full generation and heavy editing by Claude.
- Detection is less reliable on short passages because there are fewer token choices to analyze; longer passages increase confidence.
- Highly factual or constrained text (where the next token is effectively fixed) offers little opportunity for watermarking. Similarly, if Claude only makes minimal edits to human-written text, there may be too few decisions for the watermark to register.
- Light edits may not fully remove the watermark; a complete rewrite that replaces every word will eliminate it.
Special cases: code, translations, and files
- Code typically requires exact tokens, so watermarking has less effect on working code. Where arbitrary choices exist (comments or nonessential formatting), watermarking can apply but will not materially change executable code.
- Translations produced by Claude are watermarked because Claude chooses every token in the translation.
- For supported file types (for example .png, .jpg, .svg), Claude will attach a cryptographically signed content credential in the file’s metadata using the C2PA standard. This is not a visible watermark and does not alter the file content; it is a signed note indicating the file was created or processed with Claude. Any C2PA-aware tool can read that metadata.
Detection API and rollout
Anthropic plans to offer a watermark detection API; implementation details are being finalized. Because Anthropic currently lacks a robust method to apply region-specific scopes, the watermarking is being applied globally at launch to meet the EU requirements. There is a transition period in the EU for models launched before August 2, 2026; Anthropic is working to add watermarking to those earlier Claude models over the coming months.
How this differs from third-party AI detectors
Third-party AI detection services (e.g., Pangram and similar tools) use different signals because they do not possess Anthropic’s watermark key. Those detectors typically rely on linguistic patterns and stylistic “tells” that can correlate with AI text. By contrast, Anthropic’s watermark detection is a cryptographic-style check that tests whether token choices match the distribution implied by the secret key.
Legal and practical consequences
Applying a watermark does not change legal ownership, authorship, or user rights under Anthropic’s terms. It is solely a technical mechanism to estimate whether Claude was likely involved in generating or processing content.
In short: Anthropic is deploying a SynthID-Text–style, imperceptible watermark in Claude’s textual outputs to increase transparency and comply with EU requirements. The watermark is designed to be invisible to readers, inexpensive to apply, non-identifying, and detectable by parties with the appropriate key or via Anthropic’s planned detection API.
