Anthropic plans to add invisible statistical watermarks to text produced by future Claude models. The change is meant to comply with the European Union’s rules for marking AI-generated content, but Anthropic says it will apply the system globally because it cannot reliably limit the feature by region.
The watermark is not a hidden character, metadata tag or phrase tucked into every answer. It changes how Claude chooses between words that are already reasonable alternatives. When several next words would preserve the meaning, a secret key influences the random choice. Across a long passage, those small decisions create a pattern that Anthropic can test for later.
Anthropic is using a version of SynthID-Text, the watermarking method introduced by Google DeepMind. The company says its internal testing found no practical effect on creativity, readability, output quality, speed or token cost. The watermark also contains no information about the user, organization or conversation that produced the text.
There are important limits. A short passage may not contain enough word choices for a reliable result. Factual writing and code leave less room for a watermark because the correct token is often fixed. Light proofreading of human-written text may also produce too few model-selected words to register. Heavy rewriting is more likely to carry a detectable signal, while a complete rewrite can remove it.
The detector will not deliver a courtroom-style verdict either. It can estimate whether Claude was involved, but it cannot prove that a person did not write the passage or distinguish original generation from extensive editing. Anthropic says it will offer a detection API, though the implementation details have not been announced.
For images and supported files, Claude will use C2PA content credentials instead. Those are cryptographically signed records stored in file metadata, rather than patterns embedded in the content itself.
The move follows Anthropic’s signing of the EU Code of Practice on Transparency of AI-Generated Content alongside other major providers. Watermarking may give platforms and researchers a better signal than today’s style-based AI detectors, but it will not end the authorship debate. The useful question will be how strongly the signal survives editing, translation and the messy ways people mix human and machine writing.
Source: Anthropic’s explanation of Claude text watermarking.
