Since 2 August 2026, text from Claude carries an invisible statistical watermark, and files Claude generates carry signed C2PA metadata. Most of what is being said about it online is wrong in the same two ways. This page is the corrected version.
Last reviewed 13 August 2026. Anthropic has not yet published a technical specification, so this reflects the company's public statements and reporting to date.
When a language model writes, it repeatedly picks the next token from a ranked list of candidates. At most positions several candidates are near enough in probability that any of them would read naturally. A statistical watermark exploits that slack: a secret key partitions the candidates, and the sampler is nudged toward one side of the partition. Google's SynthID-Text runs a tournament between candidates on the same principle.
No single word carries the mark. The signal is the accumulated skew across hundreds of choices, which is why it needs a long passage to be readable and why changing a large share of the words destroys it. It also explains a detail Anthropic has confirmed: the model itself does not know it is being watermarked, because the biasing happens in the sampling pipeline underneath the model rather than inside it.
Error one: looking for invisible characters. Zero-width spaces and Unicode tag characters are a real watermarking technique, used by other tools, and they show up in a hex editor and die to a find-and-replace. Claude's mark is not one of these. Our inspector will happily scan your text for them, and finding none tells you nothing about Claude's mark.
Error two: treating detection as a verdict. Even once the API ships, a positive result means the text went through Claude. Someone who wrote an argument themselves and asked Claude to tidy the grammar produces marked text. Someone who had Claude write the whole thing and then rewrote a third of it may produce unmarked text. The mark tracks the tool, not the thinking.
This is a compliance deployment, not a product launch. Article 50 of the EU AI Act became enforceable on 2 August 2026 for newly launched generative systems. It requires providers to mark synthetic output in a machine-readable way so that downstream platforms and regulators can identify it. Anthropic signed the Article 50(2) Code of Practice on Transparency of AI-Generated Content, which additionally commits it to supporting third-party detection of its own marks. That commitment is what put the detection API on the roadmap.
The dates line up exactly: the obligation bites on 2 August 2026, and the watermark applies to models launched on or after 2 August 2026.
| Copy and paste | Survives. This is the primary case the design targets. |
|---|---|
| Changing fonts, formatting, HTML | Survives. The mark is in word choice, not presentation. |
| Find and replace on whitespace or punctuation | Survives. Nothing to remove. |
| Light copy-editing | Survives, with reduced confidence. |
| Heavy rewriting or paraphrase | Erased, per Anthropic's own description. |
| Round-trip translation | Erased. Translation resamples nearly every token. |
| Blending with substantial human writing | Diluted, and erased at high enough ratios. |
| Very short passages | Never readable. Statistical marks need length to separate from noise. |
None of this is a secret technique. Anthropic states the erasure conditions itself, because a watermark's limitations are part of what a transparency obligation requires it to disclose.
Article 50 does not only bind model providers. If you deploy an AI system that generates or manipulates published text, image, audio or video content, you may carry your own disclosure duties toward the people who read it. The provider-side watermark is the machine-readable layer; the deployer-side disclosure is the human-readable one, and they are separate requirements.
The practical implication for anyone running a content operation: stripping provenance from material you then publish as human-written is the wrong side of the same regulation that produced the watermark. Knowing what your pipeline does to the mark is legitimate and useful. Selling removal as a service is a different activity with a different legal profile.
Files that Claude generates also carry C2PA Content Credentials: a cryptographically signed manifest recording what produced the file and what happened to it since. Unlike the text watermark, this is verifiable by anyone today, which is why the inspector on this site works. It is also fragile in a different way. C2PA data is metadata, so re-encoding, screenshotting or uploading to a platform that rewrites files usually strips it entirely. Tamper-evident is not the same as tamper-proof, and a missing manifest is the normal condition of nearly everything on the web.
Not today, and not by any tool currently available. When the detection API ships you will be able to ask Anthropic, and the answer will be about processing rather than authorship.
No. They are statistical classifiers guessing from perplexity and sentence variation, and they predate the watermark entirely. Their false-positive rates are high enough to have caused real harm, particularly to people writing in a second language. A watermark detector and an AI detector are unrelated things.
The bias only operates between candidates the model already considered near-equivalent, so the intended effect on quality is negligible. Anthropic has not published measurements.
No. The inspector and the character scanner run in your browser via WebAssembly. The only outbound request this page can make is to a detector endpoint you configure yourself, and none is configured by default.
Google already applies SynthID-Text to Gemini output. Anthropic has said it expects others to follow, driven by the same regulation.
Anthropic's Help Center is the primary source for the watermark's existence and scope. The reporting above is secondary. Where the two ever disagree, believe the Help Center and correct this page.