What changed on 2 August 2026
Anthropic began marking the output of new Claude models at the model level. That means the marking travels with the content regardless of where it was produced — Anthropic lists the API, the Claude apps, Claude Code, Claude Cowork and Claude Tag, as well as its cloud partners at AWS, Google Cloud and Microsoft. It applies worldwide, not only to users in the EU.
The date is not a coincidence. It is the day the transparency obligations in Article 50 of the EU AI Act became applicable, which require providers of generative models to mark synthetic output in a machine-readable, detectable way.
Two mechanisms, two very different stories
Coverage of the announcement has mostly treated this as one feature. It is two, and they are not equally useful to anyone trying to check a piece of content:
Text
No published detection methodAn imperceptible watermark woven into the wording itself. Anthropic states it does not change the meaning, quality or readability of the response.
Anthropic says it is working to let users and third parties detect the mark, with technical documentation still to come. Until that lands, nobody outside Anthropic can check a piece of text for it.
Files (.png, .jpg, .svg)
Verifiable todaySigned provenance metadata following the C2PA standard, attached to the file. Tampering after signing can be detected.
C2PA is an open, published standard with open-source tooling. Anyone can read the manifest, check the signature and see who issued it — the same way OpenAI, Adobe and several camera makers already sign their output.
This is the part worth internalising. The half that made headlines — invisible watermarks in text — is the half no one outside Anthropic can act on yet. The half that got barely a mention, signed file metadata, is a working, open standard you could verify this afternoon.
A mark is not proof of authorship
Anthropic is explicit about this, and it is the single most misread part of the announcement: a detected mark signals that content may have been processed by Claude. Processed, not authored. If a student writes an essay themselves and asks Claude to fix the grammar, or a non-native speaker asks it to smooth their phrasing, the text that comes back can carry the mark.
Why this matters for schools and universities
Once the detection method is published, a watermark hit will be tempting to treat as a confession. It is not one. It cannot distinguish “Claude wrote this” from “a human wrote this and Claude proofread it” — and those two cases usually sit on opposite sides of an academic integrity policy. Any process built on watermark hits needs a human step before an accusation, the same as with probabilistic detection.
What a missing mark does not tell you
The reverse error is just as easy to make. Absence of a mark is not evidence that a human wrote something. Anthropic itself lists several ways marked content loses its mark, and several categories that were never marked to begin with:
- Older models are not covered. Marking applies to Claude models launched from 2 August 2026 onward. Anything generated before that carries nothing.
- Other providers are not covered. Open-source models, self-hosted models and other vendors mark differently or not at all. Claude is one generator among many.
- Metadata is fragile. Format conversion, re-saving, screenshots and most social platforms strip C2PA metadata from a file. The image is unchanged; the provenance is gone.
- Heavy editing degrades the text mark. Anthropic notes marks may persist through some editing — some, not all. Rewriting, paraphrasing and translation erode the signal.
- Short passages may carry too little signal. A statistical watermark needs enough text to be reliable. A two-sentence answer may not qualify.
Provenance marking is positive evidence. When a mark is present and valid, it tells you something firm. When it is absent, it tells you nothing at all — which is precisely why marking does not replace detection, and why the EU AI Act treats the two as complements rather than alternatives.
What this means in practice
Educators
Nothing actionable changes today — there is no way to check the text watermark yet. When there is, treat a hit as a reason to ask, not as a verdict.
Publishers & platforms
Start reading C2PA metadata on uploaded images now. It is a cheap, high-confidence signal that already covers Claude, OpenAI and Adobe output.
Compliance teams
Article 50(4) disclosure remains your duty as a deployer. Upstream marking does not discharge it, and unmarked content still has to be assessed.
Developers
C2PA has mature open-source tooling. The text watermark has no interface to build against yet — plan for it, but do not promise it.
Where we stand on this
We do not detect the Claude text watermark, because no one outside Anthropic can yet. When the detection documentation is published we will support it and say so plainly. Until then our text analysis remains what it has always been: a probabilistic model with published accuracy figures and stated error rates, not a provenance check.
Frequently asked questions
Does Claude watermark its text?
Yes. Since 2 August 2026, text from new Claude models carries an imperceptible watermark embedded in the wording itself. Anthropic says it does not affect meaning, quality or readability.
Can I detect the Claude watermark?
Not yet. Anthropic has said it is working to enable users and third parties to detect its watermarks and provenance metadata, but the technical documentation has not been published. Any tool claiming to detect the Claude text watermark today cannot substantiate that claim.
What is C2PA and can I check it myself?
C2PA (Coalition for Content Provenance and Authenticity) is an open standard for signed provenance metadata attached to media files. Claude attaches it to generated .png, .jpg and .svg files. Because the standard and its tooling are public, anyone can read and validate that metadata today.
Does a Claude watermark mean a human did not write the text?
No. Anthropic states that a detected mark means the content may have been processed by Claude. Text you wrote yourself and then asked Claude to edit, translate or proofread can come back marked. A mark indicates involvement, not authorship.
If there is no mark, is the content definitely human?
No. Marking only covers Claude models launched from 2 August 2026 onward. Content from earlier models, from other providers, from open-source models, or content that was heavily edited or passed through screenshots and format conversion may carry no mark at all.
Why did Anthropic start marking content?
The transparency obligations in Article 50 of the EU AI Act became applicable on 2 August 2026 and require providers to mark synthetic output in a machine-readable way. Anthropic applies the marking worldwide rather than only to EU users.