Anthropic Details Watermarking System for Claude AI Text Generation
Anthropic has published new details explaining how it plans to watermark text generated by its artificial intelligence chatbot, Claude. The move follows an announcement earlier in the week that the company would implement watermarking to comply with the European Union AI Act's Transparency Code, which mandates that AI companies use systems to make AI-generated content identifiable.
The decision has sparked active debate among Claude users. On platforms like Reddit, some users have criticized the decision, with one poster characterizing the move as a conspiracy against users, while another argued that the only reason to oppose the change is to deceive others. Additionally, Business Insider reported that dozens of users on X, formerly Twitter, claimed they were cancelling their Claude subscriptions in response to the announcement.
How the Watermarking Technology Works
In a blog post published on Friday, Anthropic outlined the technical concept behind its watermarking system. The process relies on making subtle adjustments during "low-stakes choices" when the AI selects words. For example, when choosing between similar words like "overcast" and "grey" to describe the weather, Claude can generate a specific pattern in its responses.
According to Anthropic, this pattern remains completely invisible to human readers but can be identified by anyone possessing the specific key used to encode it. The company emphasized that the technology does not degrade the performance of the chatbot, stating that watermarking does not impact the quality of Claude's output and that a watermarked response is indistinguishable from an unwatermarked one to a reader.
Anthropic confirmed it will utilize the SynthID-Text approach, which was originally outlined by the Google DeepMind team in 2024. To facilitate identification, Anthropic also plans to release a watermark detection API.
Distinction From Traditional AI Detectors
Anthropic clarified that its watermarking system is fundamentally different from existing AI detection tools, such as those offered by companies like Pangram. Traditional detectors typically scan text for stylistic "tells" or specific phrasing patterns, such as the construction "this isn't [X], it's [Y]."
The company noted that picking up on these stylistic patterns is entirely different from checking for an embedded watermark, which relies on a mathematically encoded pattern rather than stylistic analysis.
The Impact of Editing and Rewriting
Addressing questions about whether users can bypass the watermark through editing, Anthropic explained that the durability of the watermark depends on the extent of the changes. The company stated that light editing is unlikely to remove the watermark completely. However, a complete rewrite where every single word is replaced will eliminate it.
Anthropic noted that in the case of a complete rewrite, it becomes arguable whether the resulting text can still be classified as AI-generated. For scenarios where Claude is only used to proofread or lightly edit human-written text, the presence of a watermark depends on the length of the text and the depth of the edits. If the text is only lightly edited, nearly all the words remain those of the human author, leaving very little for the watermark to attach to.
Watermarking in Code and Industry Adoption
The watermarking system will behave differently when Claude is used to generate programming code. Because code must function correctly, the AI model has far less freedom to choose between different but equally valid options. Consequently, code will contain significantly less watermarking than standard text.
However, the watermark can still be applied in areas of code where arbitrary choices are available, such as within comments. Anthropic stated that this implementation will have a negligible effect on the actual code produced by the model.
Anthropic also indicated that watermarking will soon become standard across the industry. The company noted that other major AI model developers have signed the same Code of Practice under the EU AI Act and will be implementing their own watermarking systems.
