The Claude watermark, explained
Anthropic announced on August 11, 2026 that Claude embeds an invisible watermark in the text it generates. Here's what was actually announced, how a statistical watermark works, what it means for text you've already shipped — and the questions Anthropic hasn't answered yet.
The short version: every Claude model released on or after August 2, 2026 carries watermarking at the model level, with older models set to follow. The date is no accident — it is the day the EU AI Act's transparency obligations took effect, and Anthropic joins OpenAI and Google in spelling out how it will label AI output. The backlash was immediate: users don't love the idea that their employer, professor, or client could one day check whether Claude wrote their words.
What Anthropic announced
| Question | Answer |
|---|---|
| Which models? | Models released on or after Aug 2, 2026; older models to follow via extension |
| Which products? | All of them — the API, the Claude apps, Claude Code, Claude Cowork, Claude Tag |
| Text mechanism? | An invisible statistical watermark in the generated text itself |
| Files too? | Signed C2PA provenance metadata on supported files (such as .png, .jpg, .svg) |
| Does copying remove it? | No — the mark travels with copied and pasted text |
| How much editing defeats it? | Unknown — Anthropic says it "may persist through some editing" |
| Public detector? | None released so far |
| Opt-out? | None announced |
How a statistical text watermark works
There are no hidden characters to find. A statistical watermark is embedded in the word choices themselves: at each step of generation, a secret key nudges the model toward a favored subset of words. Any single sentence looks perfectly normal — but across a few hundred words, marked text leans toward the favored set far more than chance allows, and a detector holding the key can measure that lean. It's the "green list" technique we cover in depth in how AI text watermarking works.
That design has two consequences worth understanding. First, the mark survives anything that preserves the wording — copy-paste, reformatting, changing fonts, converting between file formats. Second, it degrades as the wording changes: the more heavily a passage is edited or paraphrased, the weaker the statistical signal gets. Where exactly the line sits for Claude's implementation, Anthropic hasn't said.
What the mark does — and doesn't — prove
A statistical watermark says "a Claude model generated this wording." It does not say who prompted it, why, or how much human thinking surrounds it. A heavily human-directed paragraph and low-effort spam carry the same mark. That gap between processing and authorship is fueling much of the backlash: the mark can't distinguish using Claude as a writing tool from passing off Claude's work wholesale.
It's also worth being precise about the detector question. Until a detector is public, nobody outside Anthropic can check text for the mark — but text generated today carries it permanently, so the meaningful question is not whether it can be detected today, but whether it might be detectable during the lifetime of whatever you're publishing.
Can the Claude watermark be removed?
Honestly: nobody outside Anthropic knows. Paraphrasing is the known weakness of every green-list-style watermark, and rewriting a passage heavily enough will eventually erase the signal. But "heavily enough" is doing all the work in that sentence — without access to a detector, no tool can verify that the mark is actually gone. Treat any site that promises guaranteed Claude watermark removal accordingly: it cannot check its own work.
We're researching this space — honestly
No miracle claims: today, nobody can verifiably strip a statistical watermark. If that changes — or if we build something real for the parts that can be cleaned — the people on this list hear about it first. One email, at launch, nothing else.
Should a watermark remover exist at all? Cast your vote.
Vote: should we build an AI watermark remover?Frequently asked questions
Does Claude add a watermark to its text?
Yes. Anthropic says Claude models released on or after August 2, 2026 embed an invisible statistical watermark in generated text, with older models set to follow. The mark is applied at the model level, so it shows up whether the text comes from the Claude app, the API, or Claude Code.
Can you see the Claude watermark?
No. The mark lives in the statistical pattern of word choices, not in any visible character. The text reads normally; only a matching detector can measure the lean.
Does copying and pasting remove the Claude watermark?
No — the pattern is the text, so it travels with every copy. Anthropic says the mark may also persist through some editing, though it hasn't said how much editing defeats it.
Is there a public Claude watermark detector?
Not yet. Anthropic has not released a public detector, so as of today no teacher, employer, or platform can independently check text for the Claude mark — that could change at any time.
Can the Claude watermark be removed?
Statistical watermarks weaken under heavy paraphrasing, but nobody outside Anthropic knows the threshold, and no tool can verify its own success without a detector. Anyone selling guaranteed Claude watermark removal today is guessing. Whether such a tool should exist is what this site polls.