Claude's watermark labels contact, not authorship, and your proofread draft carries it

Anthropic is adding invisible watermarks to everything Claude produces and to everything Claude merely edits; the implementation, announced in August 2026, applies the mark at the token-selection level, so it cannot distinguish text Claude authored from scratch from text a user wrote and asked Claude to polish.

Anthropic states that "marking will apply to output from supported models wherever Claude is offered", so the watermark applies worldwide and not only in the EU; the company has not explained why it chose global application over a regional one.

Generation runtimes partition model vocabularies into pseudorandom green and red token sets keyed cryptographically to preceding context tokens, and add a slight positive bias to the logits of green-listed candidates, leaving semantic coherence and inference latency intact while embedding a detectable signature.

Beginning August 2, 2026, regulatory enforcement under EU AI Act Article 50 took effect across member states, mandating that providers of general-purpose and generative AI systems mark synthetic outputs in a machine-detectable format.

All Claude models released after August 2 support the marking, and older models will be retrofitted in the coming months.

The mark is invisible and cannot be opted out of; the user has no way to see it, remove it, or check how much of their text it affects.

Within hours of deployment the watermarks-remover repository accumulated thousands of GitHub stars in less than twenty-four hours, as highlighted in Eduardo Ordax's LinkedIn industry analysis; the utility removes C2PA, EXIF and XMP metadata, purges hidden Unicode markers, and disrupts statistical logit distributions through automated localised rewriting across Markdown, PDF, DOCX and image formats.

Markings can be stripped with paraphrasing tools such as Declaude; its developer James Padolsey criticises the underlying EU regulation as arbitrary, saying the system mostly hits ordinary users while doing little against deliberate circumvention.

In an updated support article Anthropic confirmed it has signed the EU AI Act's Article 50(2) Code of Practice on Transparency of AI-Generated Content, and described compliance as embedding watermarks in text and signed provenance metadata in generated files such as PDFs, images and SVGs.

Anthropic's phrasing is that a detected mark indicates the content "may have been processed by Claude", using "processed" rather than "generated".

Anthropic says watermarking is sparser in fact-heavy passages because fewer word alternatives exist, which is a relevant caveat for legal texts that depend on precision, though there are no empirical studies on the topic yet.

Practitioners raise concerns about false positives on low-entropy outputs such as repetitive boilerplate code or structured configuration files, where limited vocabulary diversity naturally mimics green-list token selection.

Legal trade publication Artificial Lawyer's verdict is mostly relaxed, arguing the markings are harmless and that in most cases neither clients nor courts would object, but it says things get tricky if clients have explicitly banned AI use or if a judge is sceptical, because the AI contribution could be proven even when the submitted document is factually flawless.

Artificial Lawyer points out that the watermarks travel with the text: a complex contract might contain AI-marked clauses while a human wrote the rest, future contracts built on older templates carry those markings forward, and when multiple LLMs are used different watermarks could overlap in a single document; fee negotiations could shift because the AI share becomes verifiable in principle.

While proprietary API gateways can strictly enforce watermarking at runtime, open-weight ecosystems face compliance boundaries under Article 50 because downstream engineers hosting weights locally retain full control over decoding parameters, sampling temperature and custom decoding logic.

Because there is no opt-out or visibility for users, while public tooling to strip the mark is widely available, the watermark persists mainly on documents whose authors were not attempting to conceal AI use.

Anthropic's implementation operates uniformly across web surfaces, developer APIs, Claude Code and partner cloud hosting platforms without adding token overhead, latency or API pricing adjustments, and operates purely at the model sampling layer without encoding customer identity, prompt payloads or conversation metadata.

Anthropic's support article says: "When a supported Claude model generates text, it weaves an imperceptible watermark directly into the text itself. You won't see it, and it doesn't change the meaning, quality, or readability of Claude's response."

John Gruber argues in a post on Daring Fireball that no two synonyms carry exactly the same meaning, that the system sometimes boosts the probability of a worse word and lowers it for the best one, and that Anthropic's claim the difference is "imperceptible" is simply wrong.

The dev.to analysis argues that downstream systems checking for AI-generated content, such as a university plagiarism detector, a publisher's AI-content filter or a job applicant screening tool, will read watermarked text as AI-generated whether the human wrote 90% of the words or none of them.

Since marking is sparser in fact-heavy passages and detection degrades below sixty tokens, short precise high-stakes edits are the least reliably marked category, while long discursive prose is the most reliably marked.