In August 2026, Anthropic, the developer of the Claude artificial intelligence, announced that its new models will begin generating materials with an imperceptible watermark, allowing identification that the content was produced by Claude. This initiative is in compliance with the Code of Practice on Transparency of AI-Generated Content, which is part of the European Union's new Artificial Intelligence legislation.
Currently, no new Claude model has been released, and existing models do not yet have this watermark, meaning that for now, all content generated by Claude does not possess such an identifier.
Despite this, Anthropic has already clarified how this watermark will function, differentiating it from visible watermarks found in photos or videos on the internet. The Claude watermark is not inserted directly into the text, such as hidden characters; it is intrinsically linked to how the AI structures the content and persists even if the text is copied and pasted elsewhere.
To understand this mechanism, it is necessary to understand the content generation process of generative AIs, such as Claude. Basically, the AI operates by predicting words sequentially, selecting from a set of possibilities the one with the highest probability of following the previous word. An Anthropic statement exemplifies this: when considering the phrase 'The weather today was cold and…', it is highly unlikely that the next word will be 'sweetened', but it is very likely to be 'cloudy' or 'gray'.
When there are multiple options for words with similar meanings, such as 'gray' and 'cloudy', the AI makes the decision based on random numbers. The watermark alters this process, causing the selection to be guided by a key, a specific statistical code that standardizes the AI's decisions. This key influences the choice of the next word while maintaining coherence with the previous word, without compromising content quality, according to Anthropic.
This modification will not be noticed by readers. The only way to verify if the sequence of words in a text follows the Claude pattern is to have access to the corresponding key. With it, it is possible to confirm whether the text was processed by Claude. However, although the watermark allows estimating the probability of Claude's use, it is not possible to determine the exact degree of AI usage—whether it was only for translation or if the text was entirely written by it.
Translations made by Claude will receive the watermark, since the lexical choices were made by the AI. In very short texts, watermark detection will be more challenging due to the smaller sample of words. If the AI is used in a human text only for grammatical or punctuation corrections, without adding vocabulary, the mark will not appear, as the word choice was human and not statistical. Furthermore, the more the user edits the text later, the more difficult its identification becomes.
The statement adds that 'the more Claude writes, the more decisions it needs to make and the more room there is for a watermark'. In cases of strictly factual texts, where the AI has only one word option (without choosing between similar alternatives), the mark is not applied. For example, when writing the lyrics of the National Anthem, there is no possible variation, resulting in a text without a watermark.
An additional point is that images and files created by Claude will also be marked. Encrypted metadata, following the C2PA standard, will be incorporated into the files. None of these technologies will have the capacity to identify the individual or company that used the AI to create the material.
Although Anthropic has not yet disclosed all the details on how the general public can recognize the watermark in the text, the company stated that it plans to provide mechanisms for this detection.