In a blog post, Anthropic attempted to answer some basic questions about how it will watermark text generated by its chatbot Claude. Specifically, how exactly the watermarks will work, whether they can be hidden through editing, and how this will affect the code.
In a blog post, Anthropic attempted to answer some basic questions about how it will watermark text generated by its chatbot Claude. Specifically, how exactly the watermarks will work, whether they can be hidden through editing, and how this will affect the code.
Claude users have been discussing the move since the company announced last week that it was implementing watermarking to comply with the EU AI Act's Transparency Code , which requires AI companies to use systems that allow them to identify AI-generated content.
The new Anthropic post begins with a general overview of the concept of watermarking, explaining that when performing “non-critical tasks” — such as choosing between the words “cloudy” and “gray” to describe the weather — Claude can create a pattern in his responses that is “imperceptible to the reader, but is detectable in the presence of the key in which it is encoded.”
“Watermarks do not affect the quality of Claude’s answers,” the company notes. “To the reader, a response with a watermark is no different from a response without one.”
In particular, Anthropic said it will use the SynthID-Text approach developed by Google DeepMind in 2024, and also plans to release an API for watermark detection, TechCrunch reports . The company also emphasized that watermarking is fundamentally different from AI detection methods offered by companies like Pangram, which look for stylistic “markers” in text (for example, the construction “this is not [X], this is [Y]”) to detect the use of AI: “Detecting such patterns is fundamentally different from checking for the presence of a watermark.”
"Can someone simply rewrite the text to hide the watermark? Anthropic noted that it is possible, but "light editing will probably not remove the watermark completely," while "a full rewrite, where every word is replaced, will indeed do so."
"In the latter case, of course, it is debatable whether such text can be considered AI-generated at all," the company added.
As for whether the watermark will be visible on text that Claude has only proofread or edited, Anthropic said it will depend on “the length of the text and the depth of the AI edits.” If the editing was minimal, “almost all the words” would remain human-written, so “the watermark simply won’t have anything to attach itself to (or almost nothing).”
Meanwhile, program code will have less explicit labeling than plain text. The reason is that the model is required to produce working code and does not have as much freedom to choose from a multitude of equally correct synonym options.
“Meanwhile, the program code will have a less noticeable watermark than plain text, because the model has to produce working code and does not have such freedom to choose among a multitude of equally correct options.
“With this in mind, in places where there is an arbitrary choice between specific words or terms within the code — such as in comments — a watermark may be used,” Anthropic noted. “But by definition it will have little effect on the generated code itself.”
Anthropic also added that Claude won't be the only chatbot that labels text, as "other leading modelers have signed the same Code of Practice and will also implement their own watermarks."
