Anthropic has announced that all texts generated by its flagship model Claude will soon receive invisible watermarking. This is being done to comply with upcoming regulatory requirements. The company has also disclosed technical details for the first time: previously, it was only known that the watermark does not alter content and does not add hidden characters to text.
How the watermarking works
The method is based on SynthID-Text, a technology developed by Google DeepMind and published in Nature in 2024. The idea itself was proposed by Scott Aaronson back in 2022. During normal generation, the model chooses the next word using a random number. In the watermarked version, randomness is replaced by a deterministic function: it uses a secret key and several previous words. As a result, the sequence of words looks random but carries a pattern that can be verified.
At each step, all candidate words are divided into “green” and “red” lists. The model chooses from the green list slightly more often, but red words are not forbidden — they are simply less likely. The composition of the lists depends on context and the key, so there is no fixed set of words that could be memorized. Anyone who holds the key can at any time determine which list a particular word fell into at a particular position. Without the key, this is impossible.
To understand the idea, imagine a game of Monopoly where the dice are replaced by a book of digits from pi. To an outside observer, the roll seems random, but a player with that book knows all the numbers in advance. Similarly, the watermark does not change the rules of generation; it only changes the source of randomness.
In essence, this is a form of steganography: information is hidden in the choice of words themselves, not in explicit labels. The watermark adds no tokens to the text and does not increase generation cost. Readers cannot distinguish watermarked text from ordinary text either visually or by ear.
How this affects quality
Anthropic ran internal tests and found no negative impact on content, creativity, or readability. Similar results were obtained in Google DeepMind experiments with the Gemini model: user ratings (likes and dislikes) showed no statistically significant differences between watermarked and ordinary responses. In a controlled study, experts comparing texts side by side also saw no difference.
The watermark contains no identifying information: it cannot reveal who interacted with the model, in which chat, or when. At most, it allows calculating the probability that the text was generated by this particular model. For short phrases (under a few dozen words), detection accuracy drops to zero: the text needs to be long enough for the statistical pattern to become noticeable.
Notably, the model is never forced to choose rare or unnatural words. The green and red lists contain the same words the model considers in normal mode, so the text remains natural.
What’s next
The rollout of watermarking is part of a broader trend. Starting in August, regulators will require all AI service providers to label content created by their models. Anthropic is not alone: other major companies have signed similar industry commitments and are already working on their own watermarking schemes. Invisible labeling is thus becoming an industry standard, not a unique feature of Claude.
Admittedly, the technology has limitations: using the secret key yields only a probabilistic estimate of a text’s origin, not absolute certainty. Still, that is enough for regulators and provides users with an additional layer of transparency.
Bottom line: watermarking Claude’s text is a compromise between transparency, regulation, and convenience. It has virtually no impact on generation quality and makes it possible to verify the origin of text without compromising user privacy.
