Free Claude Watermark Remover & Checker for Text & Code
Paste Claude-generated text or code, see every hidden watermark character exposed in an x-ray view, and copy a clean version in one click. Free, online, no sign-up, and nothing you paste ever leaves your browser.
More: what was found, x-ray view and options
X-ray view
Your text with every hidden character exposed as a labelled chip. Hover a chip for its Unicode name.
What was found
| Code point | Character | Count | Action |
|---|
Does Claude watermark its text?
Yes, confirmed. On 11 August 2026 Anthropic announced, via an updated support page, that models released after 2 August 2026 automatically watermark the text and files they generate, with older models to follow. The mark travels with the text when copied and may persist through some editing, which places it in the statistical class, and means removing hidden characters does not affect it.
Last verified: 12 August 2026
This page tracked the Claude question through its rumour phase, and the record now has an answer. For days the claim "Claude watermarks its text" circulated without a source; Anthropic's newsroom carried nothing, and we said so. The confirmation arrived on 11 August 2026 through a support-page update, reported the same day by TechCrunch. The details matter, because the deployed reality is narrower than the rumour was.
What is confirmed, precisely
| Question | Answer at the verification date |
|---|---|
| Do Claude models watermark text? | Models released after 2 August 2026 do, automatically; Anthropic says support will extend to older models |
| Which products are covered? | The Claude API, Claude, Claude Code, Claude Cowork and Claude Tag |
| How are generated files marked? | With C2PA provenance metadata, the open standard also used for images elsewhere |
| Is the mechanism for text published? | Not in technical detail, the stated properties (travels with copied text, may survive some editing) are those of a statistical, token-level mark |
| Is there a public detection tool? | Not at the verification date |
The trigger is regulatory: the EU AI Act's transparency code took effect on 2 August 2026 and requires providers to mark AI-generated content in a machine-readable way. That is why the model-release cutoff falls on that exact date, and why Black Forest Labs, Google, Meta, Microsoft, OpenAI and Synthesia have committed to the same code.
How a mark of this class works
A statistical watermark operates at the point where the model picks the next token. At most positions in a sentence, several continuations are roughly equally good. A watermarking scheme uses a keyed pseudorandom function to nudge that choice toward one subset of the vocabulary rather than another, by an amount small enough that quality does not visibly suffer.
Read one sentence and nothing looks unusual, because nothing is unusual: every word is a word the model might legitimately have chosen. Read several hundred tokens and the accumulated preference becomes measurable to a detector holding the key. Three properties follow, and they match what Anthropic describes:
- It travels with the text. The mark is carried by which words were chosen, so copy and paste carries it along, exactly as the announcement states.
- Editing degrades it gradually. Changing a few words removes a few data points; "may persist through some editing" is the honest way to put it, and it is the phrasing Anthropic chose.
- Character-level operations do nothing. Deleting zero-width spaces, normalising whitespace or straightening quotes leaves every token choice intact.
What this means in practice
Which text carries the mark depends on the model, not the calendar alone: output from models released after 2 August 2026 is marked, output from older models is not yet, and Anthropic has said the older models will gain support. If you need to reason about a specific document, the model that produced it, not just the date it was produced, is the question to ask, and in most real situations you cannot know the model from the text.
Files are a separate channel. C2PA metadata lives in the file container and describes provenance in a checkable, signed form, and it does not survive copying the text out of the file into a chat box or plain-text editor. The two mechanisms fail in opposite ways: the text mark survives copying but not heavy rewriting; the file metadata survives nothing outside the file.
Detection claims still deserve scepticism
Confirmation of a watermark does not create public detection. Verifying a statistical mark requires the provider's key, and Anthropic has not released a detection tool at our verification date. Any third-party service claiming to detect Claude's watermark today is doing something else, usually a style classifier, with false-positive rates on human writing that make it unsafe for grading, hiring or editorial decisions.
What this tool can and cannot do for Claude text
Can
- Find and remove every hidden Unicode character in text copied out of a Claude interface, non-breaking spaces, thin spaces and soft hyphens from the chat renderer are the usual finds.
- Show you exactly where each one sits, in the x-ray view, before anything is changed.
- Demonstrate that the watermark is not made of hidden characters: clean a marked sample and the mark, living in word choice, is untouched.
Cannot
- Detect the watermark. That requires Anthropic's key, which no third party holds.
- Remove or weaken it. Character cleaning does not change token choices, this is a property of the watermark's design, not a limitation we could engineer around.
- Read or strip C2PA metadata from files. This tool works on pasted text, not file containers.
Questions about Claude's watermark
When did Claude start watermarking text?
Anthropic confirmed on 11 August 2026 that models released after 2 August 2026 watermark their output automatically, with older models to gain support later. The 2 August date is the day the EU AI Act's transparency code took effect, which is why it is the cutoff.
Can I remove Claude's watermark by cleaning hidden characters?
No. The mark is a statistical bias in which tokens the model selects across a passage; hidden characters are separate objects in the text. Removing every invisible character leaves every token choice as it was, so the watermark is unaffected.
Does Claude insert zero-width spaces into its answers?
Not as the watermarking mechanism. Invisible characters do appear in text copied from the chat interface, because the page renders typography such as non-breaking and thin spaces and copying takes them along. That is an artefact of the copy path, not the mark.
Is text from older Claude models watermarked?
Not yet, on the announced terms: the automatic marking applies to models released after 2 August 2026, and Anthropic has said support will extend to older models. Text generated before deployment carries nothing retroactively.
Can a third-party detector tell me if text came from Claude?
Not today. Statistical detection requires the provider's key, and no public detection tool exists at our verification date. Tools that guess from writing style produce false positives on ordinary human writing.
Primary sources
- TechCrunch, 11 August 2026 - the report of Anthropic's support-page announcement.
- Anthropic's Help Center - where the company published the details, and where we check before each update.
- Regulation (EU) 2024/1689, the AI Act, whose Article 50 transparency obligations drove the 2 August date.
Related reading
- How Google's SynthID marks Gemini text - the earlier deployment of the same class of mark.
- Where OpenAI stands on text watermarking - a signatory of the same code, without a deployed text mark yet.
- Hidden characters in text - what the checker above is actually looking for.
- What the EU AI Act requires of AI-generated content - the regulation behind the cutoff date.