Anthropic: Invisible Watermark, Limited Legal Scope

Anthropic now invisibly watermarks texts produced by Claude. The company claims that this signal does not affect content ownership or traceability to individuals, and that it aligns with its commitments to the European AI Act. However, the system only indicates a probable involvement of the model, a point that raises concerns about misinterpretation in education and the workplace.

Anthropic Excludes Any Impact on Ownership and Aims for the AI Act

Anthropic asserts that the presence of a watermark does not alter the ownership of content or the legal responsibility associated with it. According to the company, the marking only tests whether Claude was able to produce or process the text and does not indicate anything about the author, while not changing users' rights under its terms of use. The company explicitly links this mechanism to its commitments regarding the European AI Act. It adds that the marking cannot be traced back to a specific organization or an individual user.

However, some voices highlight that the watermark could complicate copyright claims, particularly in the case of computer code, as raised by trainer John Crickett. More broadly, the question of copyright in the age of AI remains, and the increased visibility of AI usage could rekindle these concerns.

Employers, Clients, and Universities Fear Side Effects

Users are concerned that a watermark may lead to difficulties with clients, employers, educational institutions, or in contracts when they have relied on AI as support. Freelance developer Vladislav Rajtmajer worries about a marker detected in code sent to a client, fearing challenges to authorship, while Anthropic assures that the effect of the watermark on code is "negligible." Arturo Villarroya fears complications in professional and academic contexts where AI rules are strict.

In the academic world, institutions are reintroducing in-person exams and classes without laptops, and teachers are struggling to keep up with AI-related circumventions. The watermark is presented as a potential additional detection tool, a prospect that developer Gaetan Semet supports, believing that "all" cheaters would be caught.

A Probabilistic Signal and Edge Cases Related to Editing

Anthropic warns that its system has limitations: it only indicates that a text likely involved Claude at some point. This nuance fuels fears that a positive result may be misinterpreted as a complete generation by AI. Among the edge cases, the company acknowledges that a marker may remain after a proofreading, translation, or summary performed by Claude on a text written by a person.

However, Anthropic clarifies that the watermark only applies to the words actually chosen by Claude and that, during a simple proofreading, there is very little, if anything, to which the marker can attach. Critics, such as analyst Ben Thompson, argue that attributing an authorship role to a writing tool is akin to asking a ballpoint pen to sign a work and deem it unfair that a mere grammatical correction exposes original content to a "generated by AI" label.

Data Quality and Division of Positions

Pro-labeling advocates present an industrial argument: to prevent future models from training on corpuses massively produced by previous models. Developer Donn Felker sees the risk of a "snake eating its tail," with potential degradation of quality if systems ingest their own production. Brett Levenson believes that Anthropic is at least trying to delay this scenario or to be able to identify it when it occurs.

On the transparency front, supporters argue that readers should know when a text comes from a chatbot. Aadit Sheth contends that it should be possible to identify whether a piece reflects a person's thoughts, asserting that "it's another system speaking" when AI intervenes. The initiative, now effective with the introduction of an invisible watermark in Claude's outputs, fuels a public debate marked by concerns over privacy and transparency.