Short answer: yes, as of August 2026 Claude marks its output. Here is what that actually means, in plain terms.
Does Claude watermark text now?
Yes. To meet the EU AI Act's Article 50 transparency obligations, which took effect on August 2, 2026, Anthropic began embedding an invisible, machine-readable watermark in Claude's generated text, and attaching digitally signed provenance metadata (using the open C2PA standard) to generated files like PNG, JPG, and SVG. Models launched on or after that date mark from launch; earlier models are being updated over a transition period.
What the watermark actually is
It helps to be precise about what this is and is not:
- It is a key-based, machine-readable mark, not a visible label and not the stray invisible characters you sometimes see in pasted text.
- Reliably detecting it uses Anthropic's own method, which is different from third-party AI detectors that guess from writing style.
- It is meant as a provenance signal, a way to check whether Claude may have produced or processed something.
Does it apply if I am not in the EU?
Yes. Rather than switch the marking on only for European users, Anthropic built it into the model and applied it globally. So Claude output carries the marking wherever you are and across the surfaces where Claude is offered.
Can the watermark be removed?
Anthropic itself acknowledges the marking is imperfect. For files, the mark can be lost through format conversion, re-saving, screenshots, or similar edits, so a document or image that once carried a mark may no longer show one after being changed. This is a known limitation of provenance marking across the industry, not a loophole unique to Claude, and it is why regulators expect a layered approach rather than any single technique.
It is worth separating two different things here. Removing formatting noise, the invisible characters, em dashes, and odd spacing that come with pasted AI text, is what a cleaner like textscrubr does, and it is about making text tidy and portable. That is not the same as targeting Anthropic's cryptographic watermark, and cleaning your formatting is not a reliable way to strip a model-level mark. If you need to understand your obligations in a regulated setting, treat the watermark as a compliance topic, not a formatting one.
Does a watermark mean the text is fully AI-written?
No. A detected mark means only that Claude may have processed the content. You might have supplied the ideas and wording yourself and used Claude to translate, summarize, or proofread. Equally, the absence of a mark does not prove a human wrote something, since marks can be degraded or the text may come from a tool that does not mark at all.
How is this different from an AI detector?
AI detectors do not read this watermark. They estimate, from statistical patterns and phrasing tells, whether text is machine-written, and they are unreliable, with frequent false positives. The Claude watermark is a separate, key-based mechanism that answers a narrower question, "might Claude have touched this," rather than judging authorship or quality.
The bottom line
Claude does watermark its text and files as of August 2026, worldwide, for EU AI Act compliance. It is a provenance signal with real limits: it can be degraded, it is not proof of full AI authorship, and it is distinct from both AI detectors and the everyday formatting characters a text cleaner removes. Expect Anthropic to publish more detail, including detection tools, so the specifics may change from this mid-August 2026 snapshot.