Anthropic Details Claude's Watermark and Code

Anthropic has clarified how the text generated by Claude will be watermarked, and under what conditions this marking remains detectable. The company plans to use SynthID‑Text and open a detection API, while ensuring that the output quality does not change. The framework invoked is that of the Transparency Code of the European AI Act, while user feedback continues to multiply.

Little watermarking in code and deployment by other actors

Anthropic indicates that the generated code will contain less watermarking than other types of text, as the model must produce functional code and has fewer possibilities for variation. However, the watermark may appear in parts of the code where arbitrary choices are possible, such as in comments. According to the company, this marking will have a negligible effect on the code actually produced. Anthropic also specifies that other major developers who have signed the same code of practice will implement their own watermarks.

Minor edits, complete rewrites: what the marking endures

Anthropic acknowledges that a text can be rewritten to mask the watermark: slight modifications will likely not completely remove it, while a complete rewrite where every word is replaced will erase it. In this case, the company believes it becomes debatable to qualify the text as AI-generated. For texts that are simply proofread or modified by Claude, the detectability of the watermark depends on the length and extent of the changes. If the edits are minimal, almost all the words remain those of the human author, and there is little, if anything, for the watermark to attach to.

Encoded pattern in low-stakes decisions, not a style detection

Anthropic explains that during low-stakes decisions, the model can insert an undetectable pattern for the reader but detectable by someone with the encoding key. The company distinguishes this process from AI detection methods that look for stylistic clues, such as those proposed by Pangram. According to Anthropic, spotting these writing patterns is fundamentally different from verifying a watermark.

SynthID‑Text and upcoming API, European compliance and reactions

Anthropic plans to use SynthID‑Text, an approach described in 2024 by the Google DeepMind team, and to publish a detection API. The company asserts that the quality of responses is not affected and that a watermarked text is indistinguishable from a non-watermarked text for a reader. The implementation is part of the Transparency Code of the European AI Act, which Anthropic has decided to comply with by announcing this project earlier in the week. The blog post, published on Friday, details the operation, resistance to modifications, and impact on the code. Reactions are mixed: on Reddit, some denounce a conspiracy, while others believe that refusing the watermark aims to deceive; it has been reported that dozens of users on X have canceled their subscriptions to Claude.