Signature experience · editorial specimen

The Manuscript

Claude watermarks text from future models. The mark is invisible, adds no hidden characters or extra tokens, and cannot be traced to a person, organization, or chat. It estimates whether Claude was involved; it does not prove authorship or replace a disclosure decision. Walk this labeled manuscript from draft to send, and toggle watermarked future-model text against the pre-August 2, 2026 transition.

This desk is not a detector and not a way to strip a mark. For the everyday “does Claude watermark text?” how-to, see Claude Helps. The writer’s guide on this site is watermarking and provenance.

Sit at the deskRead the writer’s guide

Writer’s desk

Draft, edited, ready to send.

The words on the page do not change when you flip the overlay. That is the honest part: Anthropic says watermarked and unwatermarked text are indistinguishable to a reader. What changes is the editorial apparatus — where a mark could live, and what a writer should do about it.

Future-model output

p. 1 · generated opening

The page Claude handed back

Editorial densityHigh · nearly every word is Claude’sIllustration of Claude-chosen words, not a detection score.

Harbor Light

First pass

The harbor had always been there, of course, waiting with a kind of quiet patience that felt almost deliberate. Mara stood at the edge of the pier and watched the light stutter across the water, as if the bay itself were trying to remember a name. She had come back to finish the letter she had started in April, the one that still sat in the drawer like an unanswered question.

Underlines mark specimen words attributed to Claude for teaching. They are not hidden characters and not a live watermark check.

Arrow keys change stage, W flips the overlay, and 1 2 3 jump when the desk has focus. Nothing is uploaded. The specimen ships with the page.

From Anthropic’s explainer

Three things a writer can cite.

What it isA mark in word choice

Claude’s text watermark uses a SynthID-Text-style method: it changes the source of randomness among equally good next words. Nothing is added to the file.

What it is notNot person-traceable

Anthropic says the watermark carries no identifying information and cannot be traced to a person, organization, or chat.

What it does not decideNot disclosure

A watermark tests whether Claude might have been involved. It says nothing about ownership or authorship, and it does not replace a disclosure decision.

Watermarking carries no identifying information and can’t be traced to a specific person, organization, or chat.
Anthropic. How Claude's text watermark works. August 14, 2026 (updated September 1, 2026). Independent publication; not affiliated with Anthropic.

What this changes for an author

The EU AI Act, as Anthropic describes its own obligation after signing the EU Code of Practice on Transparency of AI-Generated Content, requires providers serving that market to mark AI-generated text. Anthropic is applying the mark globally at launch because it does not yet have a durable way to scope it by region. That is a provider-side marking rule. It is not an author-ID regime, and it is not a substitute for the disclosure sentence in your cover letter.

Quality is the fear writers bring first. Anthropic reports no practical impact on content, creativity, or readability, no hidden characters, no extra tokens, and no added cost. The SynthID-Text paper it cites found no statistically significant quality difference when watermarking was served to a portion of Gemini traffic. You should not rewrite a scene because you think the watermark made it worse. You should rewrite a scene because it is not yet yours.

The second fear is surveillance. A watermark cannot be traced to you, your agent, your school, or a chat. A magazine that someday runs a detection API — if it is even eligible, and if Anthropic expands access beyond the private preview — would get an estimate that Claude was involved, not a name. That is also why a watermark is a poor misconduct charge: it cannot tell “Claude wrote this” from “Claude heavily edited this,” and it cannot identify another model.

The third fear is the inverse, and it is the one that will get people in trouble. Models launched before August 2, 2026 have a transition period. Proofreading human prose may leave nothing detectable even on a watermarking model, because the mark lives only in words Claude chooses. Short samples are weak. Factual passages are sparse. A complete rewrite removes the mark because it removes the machine’s choices. None of those silences are a human-authorship certificate. Keep records. See rights and ethics for the disclosure decision this desk refuses to make for you.

What this page will not do

It will not walk you through checking a file, registering for a detection API, or editing until a checker goes quiet. That everyday how-to belongs on Claude Helps. This lane is fiction, copy, and editing: what a writer owes a reader, an editor, and a future copyright examiner. Image files are a different instrument entirely — C2PA content credentials in metadata, not a text watermark — and they are covered in the watermarking guide.

Claude Writes is an independent educational publication and is not affiliated with Anthropic. Nothing here is legal advice.

FAQ

Questions the desk is for.

Does a Claude watermark identify the writer who used Claude?

No. Anthropic states that watermarking carries no identifying information and cannot be traced to a specific person, organization, or chat. Nothing in the watermark or its key would let anyone recover a user or a conversation. A mark in a submitted manuscript is not an author ID.

Should I disclose Claude help if the page might be watermarked?

Disclose according to the publisher, contest, school, client, or contract you are actually dealing with — and according to copyright-registration rules if you register. The watermark does not make that decision. Presence is not misconduct; absence is not a human-authorship certificate.

What if I drafted on a model from before August 2, 2026?

The EU law includes a transition period for Anthropic models launched before that date, and Anthropic says it is rolling watermarking onto those models over the following months. Practically, missing marks in older drafts tell you very little. Keep process records instead of arguing from silence.

Is The Manuscript a watermark detector?

No. It is an editorial desk. The underlines are a labeled illustration of where a mark can live — in words Claude chose — not a detection result. Claude Writes cannot check a file for a watermark. Anthropic’s detection API is a private preview for eligible organizations. This page is not a bypass, a stripper, or a checker.

Where is the plain “does Claude watermark text” how-to?

Everyday how-to lives on Claude Helps. This site owns the author implications: disclosure, the EU marking obligation as Anthropic describes it, and what the mark cannot prove about a person. The long writer’s guide on this site is the watermarking article.

Primary sources

AnthropicHow Claude's text watermark worksAugust 14, 2026 (updated September 1, 2026)

Official explanation of Claude text watermarking: SynthID-Text method, EU AI Act compliance, no identifying information, and what a watermark does and does not prove.

Federal RegisterCopyright Registration Guidance: Works Containing Material Generated by AIMarch 16, 2023

Official registration guidance requiring disclosure of AI-generated material and identification of human-authored contributions.

U.S. Copyright OfficeCopyright Office Releases Part 2 of Artificial Intelligence ReportJanuary 29, 2025

Official summary of the human-authorship conclusion for generative AI outputs.

Federal Trade CommissionFTC Announces Crackdown on Deceptive AI Claims and SchemesSeptember 25, 2024

FTC enforcement examples for deceptive AI claims, fake reviews, and unsupported performance promises.