Anthropic has provided more detail about its plan to watermark text generated by Claude, describing a system designed to identify AI-written material without visibly changing the response for readers. The company says the watermark will be detectable with the right cryptographic key and plans to offer a separate watermark detection API.

The clarification comes after Anthropic said earlier this week that Claude would adopt watermarking to comply with the European Union’s AI Act Transparency Code. The move has prompted debate among users concerned that employers, schools, or other institutions could use the technology to identify undisclosed Claude use.

How Claude’s watermark is designed to work

In a blog post published Friday, Anthropic said Claude will use the SynthID-Text approach developed by Google DeepMind and described in 2024. The method relies on choices a language model makes when several words or phrases would work equally well. For example, a model may have multiple reasonable ways to describe the weather. Across a response, those choices can form a statistical pattern that is not apparent to a normal reader but can be checked by someone with the relevant detection key.

Anthropic said the watermark should not affect the quality of Claude’s output. The company describes a watermarked response as indistinguishable from an unwatermarked one to a reader, which means users are unlikely to notice the system during ordinary writing or coding workflows.

The planned detection API could matter more than the watermark itself. If released as described, it would give developers, publishers, educators, and enterprise administrators a way to build verification into their own systems rather than relying solely on general-purpose AI detection tools.

Editing, proofreading, and full rewrites

Anthropic acknowledges that the watermark is not permanent in every possible version of a text. The company says light editing will probably not remove it completely, while a full rewrite that replaces every word will. That distinction makes the system more useful for identifying substantially preserved Claude output than for proving that a final document has any connection to a model.

The result may also depend on how Claude is used as an editor. If the model only proofreads a human-written passage, Anthropic says there may be little text for a watermark to attach to because most of the words came from the original author. Detection will depend on the passage’s length and how extensively Claude changes it.

That limitation is important for product teams building compliance or provenance workflows. A positive watermark result could indicate that Claude generated or substantially altered material, but it may not establish who wrote the underlying ideas, how much of the text came from a person, or whether the final version was later rewritten elsewhere.

What the system means for code

Anthropic expects the watermark to be weaker in code than in ordinary prose. Working software generally leaves a model less freedom to choose among interchangeable words or structures, reducing the number of arbitrary decisions available for a watermarking pattern.

The company says some detectable signal may still appear where code includes flexible language choices, particularly comments. It characterizes the effect on executable code as negligible by definition, suggesting the watermark is intended primarily as a provenance mechanism rather than a constraint on code generation.

For engineering organizations, that distinction could make watermarking more relevant to documentation, comments, commit messages, and generated explanations than to the functional behavior of a program. It also means a detection result may be less reliable for short snippets or highly constrained code than for longer natural-language responses.

Evidence behind Anthropic’s claims

The current details come from Anthropic’s own explanation of its planned implementation. The company’s statements about unchanged output quality, resistance to light editing, and limited impact on code are vendor claims; the source material does not provide independent test results, detection accuracy figures, false-positive rates, or a release date for the watermark detection API.

The technical foundation is not entirely unique to Anthropic. Anthropic identifies SynthID-Text, Google DeepMind’s watermarking approach, as the method it intends to use. The company also says other major model developers signed the same Code of Practice and plan to implement their own watermarks, although the source does not name those developers or describe whether their systems will be interoperable.

This is different from AI detection products that examine stylistic signals or recurring writing patterns. Anthropic specifically distinguishes watermark verification from services such as Pangram, which attempt to infer AI use from characteristics in the text. A watermark check looks for a model-created signal; a style-based detector makes a probabilistic judgment from the writing itself.

User opposition has added a market and policy dimension to the rollout. TechCrunch reported that users on Reddit criticized the move, while Business Insider reported that dozens of people on X claimed to have canceled Claude subscriptions. Those reactions are reported user statements, not evidence of a measured change in Anthropic’s customer base.

Implications for builders and enterprise buyers

For AI builders, the announcement points toward provenance becoming an API-level feature rather than a separate moderation or detection layer. Applications that generate reports, marketing copy, support responses, or internal documentation may eventually be able to attach a machine-readable check to the content they produce.

Enterprise buyers will need to decide what a watermark can legitimately prove. It may help identify content that remains close to Claude’s output, but it will not resolve mixed-authorship cases, heavily edited documents, or material copied through multiple tools. Organizations that use watermarks in employment, education, or compliance decisions will need policies for uncertain results and possible false negatives.

The approach also creates a deployment trade-off. Watermarking can support transparency requirements without exposing a visible label in every response, but verification depends on access to the right detection system. Until Anthropic publishes the API and independent evaluations, developers cannot assess its latency, cost, reliability, or compatibility with third-party content-management systems.

What to watch next

The first signal will be Anthropic’s release of the watermark detection API, including its documentation, access controls, pricing, and supported content types. Independent testing should then examine detection rates across short and long responses, paraphrasing, translation, proofreading, and adversarial rewriting.

Developers should also watch whether Anthropic publishes policy guidance for mixed human-and-AI text and how it handles code, comments, and documentation. Another open question is whether watermarks from different model providers can be recognized consistently or remain locked to each company’s implementation.

Finally, the EU’s Transparency Code and its practical enforcement will determine whether watermarking becomes a baseline requirement for major model providers or one component of a broader disclosure system.

Creati.ai perspective

Anthropic’s clarification makes Claude’s watermarking sound less like a visible label and more like a hidden provenance signal. That could be useful for large-scale verification, but only if detection tools are accurate, accessible, and interpreted as evidence of model involvement rather than definitive proof of authorship.

The most consequential test will come after the API is available. Until independent results show how the watermark behaves under ordinary editing and real production workflows, enterprises should treat it as a developing signal for governance—not a standalone compliance verdict.