Claude now has an invisible watermark: What Anthropic's new AI text system means

Anthropic has started watermarking text generated by Claude, but users will not see any obvious mark. Instead, the AI company is embedding an invisible statistical pattern into the way Claude chooses words. That pattern can later be analysed to estimate whether Claude produced the text. The move is significant because it could make AI-generated text easier to identify without changing how it looks to readers. Anthropic says the system is being introduced globally as it responds to transparency requirements under the European Union's AI Act.

How does Claude's invisible watermark work?

Claude generates text one token at a time. At many points, several words could fit naturally into a sentence. Anthropic's system uses those choices to create a statistical fingerprint. A secret key influences which suitable token Claude selects, subtly changing the distribution of words without adding visible characters or labels. A sufficiently long passage can therefore contain a pattern that a detector can test against the key. The result is a probability that Claude was involved in producing the text, rather than a simple visible stamp. Anthropic says the watermark should not noticeably affect the quality, creativity or readability of Claude's responses. The approach is related to SynthID-Text, a watermarking technique developed by Google DeepMind and described in research published in Nature.

Can the watermark identify a Claude user?

Not by itself. The watermark is designed to indicate that Claude was involved in generating or processing text. It is not a name tag containing the user's identity. It also does not prove that every word was written by AI. The system works statistically, so detection depends on having enough text and a detectable signal.

This distinction matters for students, journalists, researchers and professionals who use Claude as an editing or writing assistant. Anthropic has also acknowledged limitations. Highly precise passages, including some factual statements and code, give the model fewer opportunities to alter token choices without risking accuracy. That means the watermark may be weaker in such material.

What happens if you edit Claude's text?

The watermark is designed to survive common actions such as copying, pasting and minor editing or proofreading. But it is not an unbreakable digital seal. Major rewriting can disrupt the statistical pattern. An important limitation is also the reverse: text without a detectable watermark cannot automatically be considered human-written.

Anthropic says it plans to provide detection tools, including an API that can allow users and third parties to check for the watermark. The bigger change is clear: AI companies are moving towards systems where generated content can carry an invisible digital identity, potentially making AI provenance a normal part of writing online.