Upcoming versions of Claude will feature a text watermarking system, enabling users to assess the probability that Claude contributed to the generated text. This initiative is part of compliance with the EU AI Act, a regulation that numerous AI developers, including us, are adopting.
In this discussion, we address some common inquiries regarding our watermarking technique, its effects on Claude’s outputs, and our reasons for implementing this feature. In essence:
- Our watermarking technique does not affect the quality or content of Claude’s text. - Readers will not notice any difference between watermarked and non-watermarked text. - There are no additional characters or information embedded within the text. - The watermarking process does not require more tokens and remains cost-effective. - The watermark carries no identifying details and cannot be traced back to individual users, organizations, or conversations. - This watermarking practice is not exclusive to Claude, as since August 2, the EU mandates AI companies operating in its jurisdiction to label AI-generated content. Several other prominent AI developers have agreed to this Code of Practice and will implement their own watermarking solutions.
Watermarking is a technique that leverages the inherent randomness in the word selection process of large language models like Claude. As it generates text, Claude chooses each word from a probability distribution based on previous inputs. For instance, in the phrase “The weather today was cold and…,” the subsequent word is unlikely to be “sugary,” while options like “overcast” or “grey” are more plausible. The final choice is often determined by a random number.
Watermarking utilizes these low-stakes decisions, which appear frequently throughout the text, to create a subtle pattern within Claude’s outputs. While this pattern remains invisible to readers, it can be identified by someone with the correct encoding key. Thus, while the word choices may still seem random, they are influenced by the watermarking mechanism that changes the randomness source. This allows for a probabilistic assessment of whether a piece of text was generated by Claude.
It’s crucial to note that the model won’t show a consistent bias toward any specific words; a word like “overcast” could be chosen in one instance and “grey” in the next, depending purely on the context.
As for how watermarking impacts outputs from Claude, it does not diminish the quality in any way. A reader will not be able to distinguish between responses with or without a watermark, which contrasts with traditional watermarks found on banknotes and certain documents that can be visibly detected.
Internal tests confirm that watermarking does not alter the content, creativity level, or readability of Claude’s writing. Research from Google DeepMind on the watermarking method corroborates that there are no significant changes in user ratings between watermarked and non-watermarked outputs.
To illustrate, consider a game like Monopoly where players typically roll dice to determine movement around the board. If we switch to using the digits of pi to decide moves instead of the dice, the players' experiences remain unchanged. However, retrospectively, one could discern whether pi played a role using the sequence of moves. Similarly, watermarking does not alter the textual content but allows for verification afterward.
Claude's watermarking technique follows the SynthID-Text framework outlined by Google DeepMind, based on principles established as early as 2022. Its effectiveness, however, does have limitations. While it indicates the likelihood of Claude’s involvement, it cannot confirm whether the text originated from a human or another AI that employs a different watermarking strategy.
Watermarking also shows lower effectiveness with brief texts or factual passages, where factual correctness takes precedence over choice variability. For example, if Claude faces a clearly defined answer, such as in mathematical equations, a watermark cannot be applied.
Regarding human-edited content, since the majority of the text may belong to the original author, Claude’s moderate edits might not provide enough substance for the watermark to be detected. The more extensive the output generated by Claude, the greater the possibilities for embedding a watermark.
When it comes to code, watermarking applies only when choice exists. Strict correctness in coding limits the instances where watermarking can be used since certain tokens must be exact. Nonetheless, in areas of more subjective expression, like comments within code, watermarking can still be applied, though it will have a minimal impact on the actual executable code.
What implications does this hold for users?
Watermarking doesn’t slow down model performance or increase costs. It has a negligible effect on processing speed, and since it doesn't introduce additional tokens, the overall cost structure remains unchanged.
Importantly, the watermark is not linked to individual users, organizations, or their interactions with Claude. It strictly applies to the text produced by the model without personal identifiers.
The driving force behind watermarking is compliance with the EU AI Act. In July 2026, our company and other significant AI developers endorsed the EU Code of Practice on Transparency for AI-Generated Content, mandating these marking methods. Although we will implement watermarking globally at launch, we’re exploring more tailored approaches based on future developments.
For those interested in verifying whether a piece of text was generated by Claude, we will soon introduce a watermark detection API for users to utilize.
When it comes to image files or other supported formats, Claude will embed a content credential—a small cryptographically signed note within the file metadata—indicating Claude’s involvement in its creation or modification. This standard follows the open industry C2PA protocol, ensuring compatibility with various tools.
It is important to emphasize that a watermark is not the same as a hidden alteration; all original content remains intact. As with text, the credential merely suggests Claude's participation without disclosing any identifying information.
Although light editing might not completely remove the watermark, significant rewrites can; however, such changes may lead to a reinterpretation of the content that questions its AI-generated status.
In terms of what a watermark actually signifies, it provides an indication of Claude’s potential involvement in the content, rather than a definitive claim about authorship or creation.
Yes, translations created by Claude are also watermarked as every word is selected by the model itself.
For models released prior to August 2, 2026, the EU law offers a transition period, and we are actively working to incorporate watermarking in those versions in the ensuing months.
Lastly, it’s important to note that AI detection tools like Pangram utilize different methodologies entirely. They do not operate with our watermarking key but instead identify patterns in text indicative of AI generation, distinguishing their approach from watermark verification.
Despite these advancements, watermarking does not redefine ownership or legal responsibilities regarding content. It solely serves as a tool for assessing Claude's participation in the creation or processing of materials.
