Anthropic Has a ‘Secret’ Way to Watermark Claude AI-Generated Text: Here’s How It Works
Anthropic has detailed a text-watermarking system planned for future Claude models, offering a way to estimate whether the AI assistant was involved in writing or editing a passage.
Ni Mayank Jha··33K nabasa
Anthropic Has a ‘Secret’ Way to Watermark Claude AI-Generated Text: Here’s How It Works
Anthropic has detailed a text-watermarking system planned for future Claude models, offering a way to estimate whether the AI assistant was involved in writing or editing a passage.
The watermark will not appear as a visible label or hidden character. Instead, it will create a detectable pattern through the word choices Claude makes while generating text. Anthropic says the system will not add tokens, increase costs or carry information identifying a user, organisation or conversation.
The company plans to apply watermarking globally at launch and introduce a detection API. It has not announced a specific release date for the API or identified which Claude model will first support the system.
How Claude’s text watermark will work
Large language models generate responses by repeatedly selecting the next token from a range of possible candidates. In many sentences, several words may be similarly appropriate without materially changing the meaning.
Anthropic’s system will use these low-impact choices to leave a statistical pattern across Claude’s output. The model will continue selecting from words it would ordinarily consider, but the source of randomness behind those choices will incorporate a watermarking key.
A detector with access to the relevant key can then examine a passage and calculate the likelihood that Claude contributed to it.
Anthropic says the watermark is based on Google DeepMind’s SynthID-Text approach, which was detailed in a 2024 research paper. The technique is designed to alter the selection process without visibly changing the resulting text.
What the watermark can and cannot reveal
The watermark will indicate only the likelihood that Claude was involved in producing a passage. It will not conclusively establish that the entire text was written by Claude, distinguish writing from extensive AI editing, or identify content produced by another AI model.
It will also contain no information about the person or organisation using Claude. According to Anthropic, the watermark cannot be traced to an individual account, chat or workplace.
Detection is expected to work better with longer passages because they contain more word-selection decisions. Short responses may not provide enough information for a detector to produce a reliable result.
The watermark will also be less prominent in factual writing, mathematical answers and code, where Claude has fewer valid alternatives. The model cannot substitute an incorrect term merely to strengthen the watermark.
Editing can weaken the detection signal
Proofreading and light editing may leave little detectable evidence because most of the submitted text was chosen by the human author. The watermark can apply only to words or punctuation selected by Claude.
Translations are different because Claude chooses the complete output, giving the system more opportunities to build a detectable pattern.
Anthropic acknowledges that editing can weaken the watermark. Minor changes may leave some of the pattern intact, while a complete rewrite could remove it. The company therefore describes watermarking as an indicator of possible Claude involvement rather than proof of authorship.
Anthropic cites EU AI Act requirements
Anthropic says it is implementing the system to comply with transparency requirements under the European Union’s AI Act. The company signed the EU Code of Practice on Transparency of AI-Generated Content alongside other technology providers.
The European Commission says the marking obligations took effect on August 2, 2026. Around 190 organisations had signed the voluntary compliance code by the end of July, including Anthropic, Google, Meta, Microsoft, Mistral and OpenAI.
Anthropic intends to enable watermarking globally because it does not currently have a durable way to restrict the system by region. Models released before August 2 have a transition period, with watermarking expected to reach those models over the coming months.
Files will use content credentials instead
Claude-generated files will use a different identification method. Supported formats such as PNG, JPG and SVG will receive cryptographically signed content credentials in their metadata.
These credentials will follow the C2PA standard and indicate that Claude created or processed the file. Unlike the text watermark, this information will exist as file metadata rather than as a statistical pattern within the content.
Anthropic says the credential will not contain identifying details about the user. It is also developing a tool that will allow people to check supported files for these credentials.