Anthropic Reveals Claude's Text Watermarking Method

Anthropic has unveiled a method for embedding a machine-readable watermark in text generated by Claude that is invisible to the human eye. This measure responds to the transparency obligations of the EU AI Act, but concerns have been raised among users that traces of 'AI processing' could remain in translations, corrections, and code work.

In a statement released on the 15th of this month, Korean time, Anthropic stated that Claude's text watermark is based on Google DeepMind's SynthID-Text approach. The company explained that the watermark does not have a significant impact on output quality or content and does not create additional tokens, thus not increasing costs.

The method operates by utilizing word selection probabilities. When large language models generate sentences, they choose from several candidate words that have nearly identical meanings. In such cases, the watermark adjusts the selection criteria to leave a statistical pattern. While readers may not notice this, detection systems with the key can analyze the sentence structure to calculate the likelihood that Claude was involved.

Anthropic clarified that this marking does not identify specific individuals, organizations, or conversation content. The watermark indicates the 'possibility that Claude was involved in text processing' rather than proving who the original author is. The company also plans to provide a watermark detection API.

This measure is aligned with the EU AI Act. The European Commission announced in guidelines released on July 20 that the transparency obligations of Article 50 of the AI Act will take effect from August 2, 2026. This provision requires machine-readable markings on synthetic audio, images, video, and text outputs generated by AI.

SynthID is a watermarking technology designed to embed hard-to-detect markings in AI-generated products, allowing post-detection systems to identify them. In text, it does not hide separate characters within sentences but embeds patterns in the token selection process. Therefore, it is nearly indistinguishable from regular sentences to those reading the original text.

Google DeepMind has described SynthID as a technology that adjusts token probabilities during the text generation process to embed watermarks. OpenAI also announced on May 19 that it applied C2PA and Google SynthID-based watermarking to images in its content source attribution update, and on July 31, it expanded this to audio. There is a confirmed trend among major AI companies to broaden the source attribution systems for their generated products.

Previous reports from our publication have confirmed that Claude-generated text includes embedded watermarks, and supporting files carry source metadata in a digital signature format. The C2PA method attached to files is closer to leaving source records in the metadata of the output, while the text watermark analyzes the selection patterns of the content itself, which is a key difference.

Reactions have varied. Those advocating for enhanced transparency believe that a technical mechanism is needed to distinguish AI-generated products from human-written ones. The argument is that without a source attribution system, the verification costs could increase in situations where AI outputs mix with search, educational, business documents, and creative works.

Opponents worry that the watermark may not effectively distinguish between 'writing' and 'assisting' in practical work. Business Insider reported that some Claude users canceled their subscriptions due to the watermark. Users raised concerns that even when AI only partially contributed to tasks like code writing, translation, correction, or drafting, traces could remain.

Anthropic explained that in such cases, the watermark does not change copyright or ownership. The detection of the watermark is merely a signal indicating the possibility that Claude was involved in text processing, and does not lead to the conclusion that the entire document was written by Claude. There are calls for the detection results to be interpreted alongside the context of document creation in professional settings.

There are also technical limitations. Anthropic stated that detection may not work well in cases of short samples, fact-based sentences, or corrections where the model has fewer word choices. They also acknowledged that complete rewriting could weaken or remove the watermark.

This indicates that the watermark is a supplementary means of source verification, not a standalone judgment tool. In outputs like lengthy creative texts where the model has many options, patterns are more likely to remain, but in texts with many numbers, proper nouns, or fixed phrases, the signals may weaken. If detection results are used as direct grounds for disciplinary actions or copyright judgments, disputes could arise.

The impact on domestic users is also significant. In cases where Claude is used for translation, correction, development assistance, or corporate document creation, there may be a need to explain whether the final output is an AI-generated product or an AI-assisted work. This is particularly important for organizations that scrutinize the creation process, such as schools, media, and corporate security departments, which will need to establish separate detection criteria and clarification procedures.

This issue boils down to the balance between identifying AI-generated products and user trust. While transparency obligations have been strengthened, how the watermark will effectively distinguish between writing and assistance in real-world scenarios remains an open question. Anthropic plans to provide a watermark detection API, and the transparency obligations of the EU AI Act are already in effect from August 2, 2026.