Anthropic unveils its watermark: limited detection and announced API

Anthropic explains how it marks texts produced by Claude without inserting hidden characters, with detection reserved for its own models. The company promises a verification API and a global rollout aligned with the EU AI Act, while acknowledging significant limitations, from code to complete rewrites.

Detection is specific to Claude and varies by text type

According to Anthropic, the verification key only addresses the likelihood that a passage was partially written by Claude and does not pertain to other AI systems. Longer texts are easier to flag, while factual passages will be less marked. For code, the company indicates that the watermark will be applied less frequently due to precision requirements, as an inaccurate term could prevent software from functioning. Anthropic warns that a slight edit is likely not enough to completely remove the mark, whereas a complete rewrite replaces every word and erases it to the point where it becomes debatable to still classify the text as AI-generated. The presence of the watermark indicates that Claude has processed the content but does not alter, according to its terms, the user's rights or ownership of the text.

Concerned subscribers, reported cancellations, but no increase according to the publisher

Before the blog's publication, clients expressed concerns about the implications of the marking. Some even indicated they had canceled their subscription to Claude for this reason, while Anthropic claims not to observe an upward trend in cancellations since the announcement. Users are particularly questioning authorship attribution when the watermark is present. Meanwhile, OpenAI did not immediately respond to a request for comment, but a support page updated two weeks ago mentions a plan to also add watermarks to texts.

A watermark through word choice, imperceptible and verifiable via API

Anthropic describes a marking derived from word choice decisions made by its models, part of which is random. The company presents the process as another random number generator, whose consistency is detectable with a key. This key is not published at this stage, but an API is announced to allow users and third parties to verify the presence of the mark. The watermark does not incorporate hidden characters, invisible fonts, or texts and relies solely on vocabulary, in a way that is intended to remain imperceptible to the reader. Anthropic refers to a 2024 article from Google DeepMind, whose technique inspires the one now applied by Claude.

Compliance with the EU AI Act and announced global rollout

Anthropic applies a watermark to texts produced by its AI and links it to the transparency requirements of European regulation. Following an update to its support page this week, the company published an article on Friday detailing the approach and its limitations, asserting that other AI providers will need to follow similar obligations. It plans to activate the marking globally at launch, lacking a robust geographical restriction solution, prioritizing new models first and then older ones in the coming months.