Does Claude watermark its text?

Anthropic's help centre states that Claude embeds an invisible watermark in generated text that travels with it when copied, and attaches signed C2PA provenance metadata to supported image files. Detection of the text watermark is offered through a Watermark Detection API in private preview for eligible organisations. Syft does not have access to that API and does not read the watermark.

Can Syft detect Claude-written text?

Syft Detect reports where essay-style writing carries AI-like signals, and marks the passages behind the band. It measures features of the writing rather than the model that produced it: Claude output that is formulaic and uniform reads as AI-like, while output that has been edited or written in an unusual register may not. Syft has not published a measured detection rate for Claude and does not claim one.

Is it true that Claude never shows up on AI detectors?

That claim circulates widely, usually from single examples. Detectors differ, Claude versions differ, and prompts differ. Independent benchmarks show that statistical detectors lose accuracy on generators and styles they were not built around, which applies to Claude as to any model. Treat both 'always' and 'never' as unsupported.

What would a reliable answer need?

A dated test with a named Claude version, a documented prompt and decoding setup, a human-written control set from the same domain, and results reported as recall at a fixed false-positive rate. Syft will publish its own figure only in that form.

Can AI detectors detect Claude?

Syft Detect reads essay-style prose and reports where the writing carries AI-like signals, whichever assistant produced it. Here is what is actually known about Claude, what Anthropic itself does to mark its output, and a check you can run on your own text.

Published 28 September 2026Syft

The short answer

Three separate things are true at once.

Anthropic marks Claude's output

OBSERVEDClaude embeds an invisible watermark in generated text and attaches signed provenance to supported image files, per Anthropic's help centre. Reading the text watermark requires Anthropic's detection API, which is in private preview.

Syft Detect reads the writing, not the model

INFERREDSyft Detect measures features of the prose and marks where the signal concentrates. Formulaic, uniform Claude output reads as AI-like; edited or unusually styled output may not. That is true of every generator.

No published Syft figure for Claude

UNKNOWNSyft has not measured and published a detection rate for Claude output. This page will carry one only when a dated, controlled measurement exists.

Check text

Run Syft's check on text you think Claude wrote.

Syft Detect marks the passages that carry AI-writing signals and names the pattern behind each one. It is the same reading for any generator; what changes is the writing.

Your text

0 words · 0 characters

Free analysis. A free account is needed to view and save the result. Your text is encrypted while it waits and is deleted if you don’t open it.

Syft AI

Ready when you are.

Paste at least 120 words for a meaningful estimate.

The same Syft AI analysis, on every plan
Passage-level evidence, not just a number
Free account to view and keep the result

What Anthropic says

Claude marks its own output. That is the strongest signal, and it is not one Syft reads.

Anthropic documents its marking in a help-centre article. The points below are taken from it; the article is the source, and it changes.

OBSERVEDGenerated text carries an invisible watermark that travels with the text when it is copied and pasted elsewhere; supported image files carry signed C2PA provenance metadata.

OBSERVEDA free content checker verifies file credentials; a Watermark Detection API is in private preview for eligible organisations such as educators, researchers, media and regulators.

OBSERVEDAnthropic states the marks are a signal, not proof: detection is “not fully conclusive”, the absence of a mark does not mean content was not AI-generated, and edited, very short or older-model content may carry no detectable mark.

INFERREDFor a teacher or editor, this means the strongest signal for Claude text exists but is not generally available, and even it is framed by its maker as evidence rather than proof. What Syft Detect contributes is a reading of the writing itself, with the passages shown.

What a statistical check sees

Writing features, whichever model produced them.

Syft Detect reads passage by passage and identifies how strongly the writing carries patterns associated with machine-generated text. It has no notion of “Claude”. A passage reads as AI-like when its features do, and Claude can produce prose that does and prose that does not, depending on the prompt, the version and any editing afterwards.

So the honest answer to “can it detect Claude” is: often, on stiff generic output, and the marked passages show you where. Less reliably as the output is edited or the register changes — independent benchmarks report that detectors lose accuracy on generators and styles they were not tuned for, and on paraphrased text — and never as proof of authorship.

The widely repeated line that Claude “doesn’t show up” on detectors comes from individual examples. It is neither confirmed nor refuted by anything Syft has measured.

References

01How Claude marks AI-generated content Anthropic Help Center · Anthropic · 2026
02RAID: A Shared Benchmark for Robust Evaluation of Machine-Generated Text Detectors Dugan et al., 2024 · ACL · 2024
03Testing of detection tools for AI-generated text Weber-Wulff et al., 2023 · International Journal for Educational Integrity · 2023

Model and version context

What changes, and what to check the date against.

Detection questions are version-specific. A page like this decays as models change, which is why it carries an updated date and says what it does not know.

OBSERVEDAnthropic states that marking applies to supported Claude models across its products, with rollout completing by 2 December 2026 (per the help article at the time this page was written).

UNKNOWNSyft has not yet published a measured accuracy or false-positive rate for Claude output. No figure is stated here because no approved measurement of the production engine exists. When one does, it will appear in this place with its engine version, method, sample, date and limitations.

Questions

About detecting Claude.

Does Claude watermark its text?: Anthropic's help centre states that Claude embeds an invisible watermark in generated text that travels with it when copied, and attaches signed C2PA provenance metadata to supported image files. Detection of the text watermark is offered through a Watermark Detection API in private preview for eligible organisations. Syft does not have access to that API and does not read the watermark.
Can Syft detect Claude-written text?: Syft Detect reports where essay-style writing carries AI-like signals, and marks the passages behind the band. It measures features of the writing rather than the model that produced it: Claude output that is formulaic and uniform reads as AI-like, while output that has been edited or written in an unusual register may not. Syft has not published a measured detection rate for Claude and does not claim one.
Is it true that Claude never shows up on AI detectors?: That claim circulates widely, usually from single examples. Detectors differ, Claude versions differ, and prompts differ. Independent benchmarks show that statistical detectors lose accuracy on generators and styles they were not built around, which applies to Claude as to any model. Treat both 'always' and 'never' as unsupported.
What would a reliable answer need?: A dated test with a named Claude version, a documented prompt and decoding setup, a human-written control set from the same domain, and results reported as recall at a fixed false-positive rate. Syft will publish its own figure only in that form.

Check text How Syft measures AI-like writing