Anthropic Explains How Claude’s New Text Watermarks Will Work

Anthropic has provided more details on how it plans to watermark Claude-generated text, explaining how the system will identify AI-generated writing while remaining invisible to readers.

The company plans to use an approach based on SynthID-Text, a watermarking technique developed by Google DeepMind. Anthropic is introducing the system as part of its efforts to meet transparency requirements under the EU AI Act.

How Claude’s text watermark will work

Rather than adding visible labels or hidden characters, the watermark will create statistical patterns based on the choices Claude makes while generating text. When several equally appropriate words could be used, the model can favour particular choices in a way that forms a detectable pattern.

Anthropic said the process should not affect the quality or appearance of Claude’s responses. To a person reading the text, a watermarked response should look the same as one without a watermark.

The company also plans to release a watermark-detection API to check whether text contains the statistical signal.

Anthropic stressed that watermark detection differs from conventional AI detectors, which attempt to identify machine-generated writing by analysing language patterns or stylistic characteristics. A watermark instead looks for a signal deliberately inserted during generation.

Editing can weaken or remove the watermark.

The watermark is not designed to survive every possible modification. Anthropic said light editing is unlikely to eliminate the signal, while extensively rewriting the text and replacing essentially every word can remove it.

Whether a watermark appears in text that Claude has only proofread or lightly edited will depend on how much of the final wording was actually generated by the model. If most of the text remains human-written, there may be little AI-generated material available for the watermark to encode.

Code will carry a weaker watermark

Generated code presents another limitation, as Claude often offers fewer choices when producing software that must function correctly. That leaves less flexibility for inserting the statistical patterns used by the watermark.

Anthropic said arbitrary elements, such as comments or certain naming choices, may still provide opportunities for watermarking, but the company expects the effect on functional code itself to be negligible.

Anthropic also said other major AI developers that signed the relevant EU Code of Practice are expected to introduce their own approaches to identifying AI-generated material.

The additional explanation makes clear that Claude’s watermark is intended as a machine-detectable provenance signal rather than a visible label. It can survive some editing, but Anthropic does not claim that it will remain detectable after a complete rewrite.