How Claude Marks AI-Generated Content

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:

This is the part worth internalising. The half that made headlines — invisible watermarks in text — is the half almost nobody can act on, because the key is private and the detection API that uses it is gated to a short list of organisations. The half that got barely a mention, signed file metadata, is a working, open standard you could verify this afternoon.

What Anthropic disclosed on 14 August

Twelve days after switching the marking on, Anthropic named the method: the text watermark is a version of SynthID-Text, the scheme Google DeepMind published in Nature in 2024. At each step the model has several near-equivalent words available, and a secret key nudges which one it picks. The result reads normally, but the sequence of choices carries a statistical pattern that anyone holding the key can test for — and, as Anthropic puts it, only they hold it.

That closes one question and opens another. The mechanism is no longer a black box; it is a peer-reviewed method whose properties are documented, and it is the same family of technique behind Google's SynthID. But knowing how a lock works does not open it. Verification still needs the key — and on 1 September 2026 Anthropic began handing out limited access to it.

What changed on 1 September: a detection API, in private preview

Anthropic has released the detection API — but not to the public. It is in private preview, and the eligibility list reads like the EU AI Act's own enforcement chain: regulators, law enforcement, media, fact-checkers, independent researchers, educational organisations and EU civil society groups, plus enterprises that are themselves obliged to verify watermarking under the Act. Organisations can register interest; Anthropic says it plans to expand access over time.

Note what the API actually answers. In Anthropic's framing the question is “what is the likelihood this was partly written by Claude?” — a probability, and a probability about partial involvement. It does not confirm that a text was human-written, it cannot tell you whether a different AI produced it, and Anthropic states it does not work well on short samples, where there are fewer word choices to carry the signal.

The limits, in Anthropic's own words

The announcement is unusually candid about what the watermark cannot do. These are not our caveats bolted onto someone else's claim — they are the vendor's:

Read together, those points describe a signal that is useful at scale and unreliable on the single document — which is exactly the case an educator or an editor actually has in front of them.

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.

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

Frequently asked questions