Claude · The myth-buster
Does Claude watermark text? Yes, and here is what that means
Short version: yes, Claude marks its text now, and no, that doesn't make detection proof.
Does Claude watermark text? Yes. Models launched in the EU on or after August 2, 2026 carry imperceptible text watermarks. The mark is applied at the model level. It is present no matter which Claude product the text comes from. It travels with copy and paste, and it may persist through some editing. Older models are being updated. Anthropic says detection tooling is still forthcoming. There is no off switch. The marks have limits. Heavy editing, paraphrasing, translation, and very short passages can weaken or erase a mark. A detected mark is not conclusive proof.
Why this page exists
People search this more than any other question here. The results are stale, and many still say Claude does not watermark text. The answer changed in August 2026. Searchers need that correction stated plainly, with the date. So we state it here.
The tells that show up here
Not every AI tell appears everywhere. These are the ones that cluster on this surface, and what it is about the format that produces them.
- vague-attribution
- Claims about what a vendor does need a named source and a date. 'Experts say' with no name is the pattern to strike.
- not-x-its-y
- The 'not X, it's Y' pivot shows up in stale and overcorrected content alike. This page answers flatly, without that pivot.
- chatbot-artifact
- Validation openers that thank or praise the reader before answering. Strike them from the draft.
Before and after
A generic AI draft on the left. On the right, the real output of running that draft through underslop, checked so it adds no name or number the draft did not already carry. The voice is The Steady Hand.
Great question, and it's one more people are asking every day. The short answer is no — as of right now, Anthropic has not deployed a public, verifiable text watermark in Claude's output. It's not just a matter of technology; it's a testament to the company's commitment to user freedom in an ever-evolving regulatory landscape.
Yes. Claude text carries watermarks now. Models launched in the EU on or after August 2, 2026 mark text at the model level. The mark is imperceptible. It travels with copy and paste, and it may persist through some editing. Generated files get a second mark. Files like .svg, .png, and .jpg carry signed C2PA metadata. The marks have limits. Heavy editing, paraphrasing, translation, and screenshots can weaken a mark. A detected mark is a signal, not conclusive proof. Detection tooling is not published yet. There is no off switch.
How to do it by hand
- State the current date August 2026: the answer changed. EU-launched models began carrying machine-readable marks at launch.
- Describe the two mark types Cover the two marks separately. The text watermark sits at the model level. The C2PA metadata sits on generated files.
- Quote the limits from Anthropic Use the vendor's own limitation list. Heavy editing, paraphrase, translation, screenshots, and format conversion can still cost a mark.
- Say what detection exists Anthropic’s detection tooling isn't published yet. The detectors available now work statistically. They don't read watermarks.
Questions
Does Claude leave watermarks in text?
Yes. Models launched in the EU on or after August 2, 2026 carry embedded text watermarks. The mark sits at the model level, and older models are being updated.
Which AI models watermark text in 2026?
Anthropic now marks Claude text in the EU, from August 2026. OpenAI says text provenance comes later. Google's verification is live in its Gemini app.
Will the EU AI Act force Anthropic to watermark?
The EU AI Act's transparency rules take effect August 2, 2026. Under the Code of Practice, outputs should be machine-readable and clearly detectable as artificially generated, where feasible.
What is the difference between a watermark and a detectable style?
A watermark is a deliberate signal the provider embeds in the text. A detector's flag is a statistical judgment about how predictable the text looks.
Can I turn the watermark off?
As of August 2026, Anthropic lists no off switch, setting, or admin control for marking. Marking is part of how new models ship. There is nothing to switch off.
Claude marks are real and the coverage is still settling. underslop is the edit, not a remover: it edits an AI draft so it reads like you wrote it, and it makes no promise about what any detector will report. Your text is held for 0 seconds.