[DigitalToday reporter Hyunwoo Chu] Anthropic has disclosed how an “invisible” electronic watermark inserted into text generated by its AI chatbot Claude works, Japanese IT outlet ITmedia reported on Aug. 16.

According to the explanation released on Aug. 14 (local time), the approach is based on Google DeepMind’s watermarking technology SynthID-Text. It replaces the random-number generation method used for word selection with a secret key and the immediately preceding context to leave a statistical pattern.

Anthropic said it will embed watermarks sequentially in text generated by Claude after signing a code of practice linked to the EU AI Act’s transparency obligations. The scope is not limited to the EU and extends across models and products worldwide, including South Korea. Anthropic explained, "There is not yet a reliable way to limit the scope by region, so at launch we apply the watermark globally." About 190 organisations, including Anthropic, signed the code of practice, and other major AI developers also plan to implement their own watermarks.

■ Nothing added to text... no impact on speed or price

Large language models generate text in word, or token, units. Anthropic cited as an example that after a phrase like "Today the weather is cold and," low-risk options such as "overcast" and "cloudy" that are acceptable either way in meaning appear repeatedly throughout a sentence. Such choices are usually decided by random numbers, but with a watermark applied, the model generates the random numbers based on a secret key and the last few words, leaving a pattern that only the party holding the key can identify.

Anthropic said it does not steer the model to overuse certain words or to pick words it does not normally use. It said nothing is added to the text and there are no hidden characters, so no separate tokens are needed for watermarking, prices do not rise and the impact on speed is minimal. The watermark or key does not include information that can identify a user, organisation or conversation.

■ "A full rewrite makes the watermark disappear"

Detection works by using the secret key to verify word sequences via a dedicated API that Anthropic will provide soon, measuring the probability of Claude’s involvement. Still, sentences with limited freedom in word choice leave little room for a watermark. Factual statements with only one correct answer such as "Isaac Newton’s most famous book is 'Principia...'" and formulas such as "2+2=" are excluded. Programming code, which malfunctions if words are changed, applies the watermark only to comment sections where arbitrary choice is possible. By contrast, translations in which Claude selects every word anew include the watermark.

There are also limits. Anthropic explained, "Light editing does not completely erase the watermark, but it disappears after a full rewrite that changes every word." It added that in such cases, whether the result can itself be called AI-written becomes a point of debate. The watermark only indicates the possibility that Claude was involved in some form and cannot distinguish between "Claude wrote it directly" and "Claude heavily edited it." It also cannot prove that a human wrote it or determine whether a rival AI model generated it. Accuracy drops with shorter text and rises as the text gets longer.

■ Images use a C2PA metadata approach

When Claude generates image files such as PNG, JPG and SVG, it uses a different method from text. It adds a cryptographic signature record to file metadata indicating the content was produced or processed by Claude, following the content provenance standard C2PA. This is the same standard used by camera makers and photo-editing software and can be checked with C2PA-compatible tools. Anthropic plans to provide its own verification tool as well.

Existing models released before Aug. 2, when application of Article 50 of the EU AI Act (transparency obligations) began, are subject to a legal transitional measure. Watermark additions will be applied sequentially over the coming months.