Claude Watermark Remover

Claude started watermarking its own writing on 2 August 2026

Claude has watermarked its own text since 2 August 2026. Stripping invisible characters is not enough — the mark lives in the word choices. The rollout applies to Claude models launched on or after that date; older models have a separate transition. Do not infer the marking status of a particular passage from its date alone.

This page separates three problems that often share a name. You can clean characters, revise sentences, or inspect a file credential. Those actions produce different results. Start by deciding which problem your draft actually has, rather than treating a cleaner's character count as evidence about its origin.

How an AI text watermark works, in plain English

A statistical watermark works at token selection. A secret key acts as a seed for repeatable randomness. A simplified picture divides plausible next words into groups and favors one group slightly. Context changes the choices, so there is no fixed list of suspicious words to delete. A reader sees ordinary writing; a holder of the correct key can test the pattern statistically.

The practical consequence is straightforward: deleting zero-width characters, changing fonts, copying and pasting, or doing ordinary find-and-replace does not remove this statistical mark. Changing the sequence of chosen words through rewriting can affect the pattern. Formatting cleanup and rewriting therefore need separate labels, separate expectations, and separate review.

Can anyone check whether your text is watermarked? Not yet

There is no publicly available tool here that can verify whether Claude's statistical text watermark has been cleared. Anthropic's detection API is in private preview for eligible organizations, rather than an open public detector. This site has neither the relevant key nor access to that verification service.

We do not display a detection score, a remaining-watermark percentage, or a certificate. A count of deleted characters means exactly that: those characters were present and were removed. A readable output is useful, but readability is not verification. Keep that distinction when comparing tools, interpreting a report, or deciding whether an edited draft is ready to submit.

Three different marks get called "the Claude watermark"

A PNG or JPG credential, a PDF's supported provenance metadata, and a sentence's token pattern are not interchangeable. File support varies by creator and viewer. A file extension alone is not proof that a Claude credential exists. Always inspect the actual file with a compatible viewer.

What this tool actually does

Light performs only selected character operations in your browser. Standard and Deep request wording changes through our server algorithm after you confirm the processing notice. Standard requests restrained edits; Deep requests more substantial sentence restructuring.

The prompt and post-processing protect numeric values, dates, proper-name candidates and your allowlist, URLs, code blocks and quoted passages. Missing protected items stop a rewrite and appear in the preservation report. Name candidates are mechanically extracted; the system does not claim exhaustive recognition of every entity.

Selected character operations have their own local preview and diff. Server wording is checked against your original, preserving quoted and code content. The two branches are labeled separately. Zero-width operations can affect joining scripts and emoji; choose those switches deliberately. Neither branch supplies a Claude provenance verdict.

How to remove it in three steps

1. Paste and inspect your draft. Paste text you are entitled to edit. Open the character view to see invisible Unicode. A highlighted square identifies a character, not a token watermark.
2. Choose a mode and clean text. Select Light for precise cleanup, Standard for spacing, or Deep for supported sentence reordering. Choose the punctuation and Unicode switches deliberately, then press Clean text.
3. Review and copy the result. Compare the result with your source, especially names, numbers, quotations, and clause meaning. Copy only after review. No step verifies that the statistical watermark is gone.

For formatting issues, these steps remove or replace the selected characters. For statistical marking, they are only an editing workflow with limited rewriting. Save your original separately before experimenting, particularly when the draft contains technical instructions or commitments.

Does rewriting actually weaken the watermark?

Rewriting changes the sequence in which words occur, which can disrupt a pattern tied to those choices. That is a mechanism, not a measured result for your draft. A handful of clause moves may leave most of the passage untouched. Our controlled own-key bench cannot tell you what Anthropic’s detector would return, and the report makes no such claim.

Do not judge effectiveness by how different a paragraph looks. Changed punctuation can look dramatic while leaving the words intact. Conversely, a restrained edit can alter emphasis in a way that matters to a reader. Evaluate the actual edit against your writing goal. If you need a fresh expression of an idea, write from your own notes and check your facts, rather than treating cosmetic variation as proof.

Claude marks files too, and that mark you can actually verify

C2PA Content Credentials are cryptographically signed provenance metadata. A compatible viewer can inspect an embedded credential and its validation state. Anthropic provides a browser-based checker for supported files, including PNG and JPG. That file checker does not check a text watermark. PDF provenance depends on file and viewer support; we do not claim every Claude PDF receives the same credential.

Embedded C2PA metadata can be stripped, but stripping removes provenance information, not necessarily every way an asset might be associated with its history. Absence of an embedded credential does not establish human authorship. This page handles pasted text only: it neither uploads files nor strips their metadata. Keep the original file if its provenance matters to a collaborator or reviewer.

When Claude marks writing that was yours to begin with

A draft can start with you and still be edited by a model. Separate the original composition from the changes you asked for. Keep a dated original, a version history, and notes about which parts were revised. These are practical records of your process, rather than a claim about what any detector might decide.

For an editor, employer, or teacher, explain your workflow in the terms their policy asks for. “I wrote the draft and used an assistant for punctuation” describes a different process from generating a complete passage. Cleaning invisible characters does not change that history. If authorship becomes a dispute, a character report from this site is not evidence that the wording was written entirely by a person.

Should you remove it? A straight answer

Reasonable uses

Cleaning your own drafts, removing unwanted invisible characters for layout, and considering privacy before sharing are reasonable editing goals. A copied draft may have awkward spacing or punctuation that does not fit your publication. Make those changes because they improve the document. Preserve an original when collaboration, archival requirements, or a review process depends on an accurate history.

Things this does not make okay

Do not present AI-assisted work as independently authored where course assignments, contracts, regulatory filings, or other rules require disclosure. This tool does not override an institution's policy, excuse a false statement, or establish compliance. Check the applicable requirements before submitting. In mainland China, you must independently confirm local rules concerning removal of AI markings; this site does not determine legality for you.

If a requirement is unclear, ask the receiving institution what it expects. A formatting tool cannot answer a policy question about the provenance of your work.

Which rewrite mode should you use?

Use Light when spacing is deliberate: poetry, quoted text, alignment-sensitive notes, and technical examples deserve minimal intervention. Turn off any switch that changes a meaningful feature. Use Standard for everyday prose with stray whitespace or inconsistent punctuation, when you want the same words in a cleaner presentation.

Try Deep only if your goal includes revising sentence structure and you can review the wording. It requires confirmation that your wording will be sent to our server and its third-party processing service. A returned passage may still be too close to its source; review the actual diff, verbatim span and preservation report. Choose another editing workflow when the draft needs substantial vocabulary changes or a complete new expression.

For legal wording, quotations, code, names, measurements, and multilingual text, start with Light and compare carefully. Mode names describe editing intensity here, not levels of demonstrated watermark removal.

What happens to what you paste

Light mode never leaves your browser.

Standard and Deep rewrite your wording and are processed on our servers. The server algorithm uses a third-party processing service. This application does not save request text, response text, or text-bearing logs. We cannot promise that processing or hosting providers have no retention; review their policies before sending sensitive text.

The server does not persist visitor input or output in an application database, cache, analytics event or text-bearing log. It holds a request while processing and responding, rather than promising instantaneous physical memory erasure. The browser retains the working draft so a failed run can be retried; refreshing or closing the page discards it.

The controlled bench is a separate server workflow. A prompt supplies context for a newly generated test sample carrying our own-key signal. It does not retroactively mark the original words. This distinction matters when interpreting before-and-after measurements.

Copy writes the result to your clipboard, which operating systems or extensions may retain. Optional analytics sends only counts, numeric costs, booleans and allowlisted enums. No visitor text, character preview, protected strings or provider payload is included. Read the same processing notice on the privacy page before using server modes.

FAQ

Sources

Primary references, checked 6 October 2026. Product behavior and availability can change. Read the current source when a policy or verification decision matters.

• Anthropic: How Claude’s text watermark works ↗
• Anthropic: Check if files were made with Claude ↗
• C2PA: Content Credentials technical specification ↗