Does Anthropic really watermark Claude's text?

Yes, since 11 August 2026. Every Claude model launched on or after 2 August 2026 embeds an imperceptible mark in generated text, applied at the model level so it is present on every Claude surface, and applied worldwide rather than only in the EU. The driver is EU AI Act Article 50, which became enforceable on 2 August 2026. Anthropic states the mark travels through copy-paste and may persist through some editing.

Can the watermark actually be removed?

Anthropic's own documentation says the mark may not survive heavy editing, paraphrasing, translation, mixing with other content, or very short passages. Paraphrasing is on that list, and the rewrite here is a paraphrase pass. But Anthropic has not published its detection mechanism, so nobody outside Anthropic - including us - can currently verify whether a mark is present or gone. We will not claim a result we cannot demonstrate. When a detector is published we will test against it and publish what we find.

What does a detected mark actually prove?

Less than people assume. Anthropic states a mark indicates content may have been processed by Claude - not that Claude authored it. Text you wrote yourself and asked Claude to proofread carries the same mark as text Claude drafted from scratch. The mark cannot distinguish those two cases.

What are the two layers you clean?

Character-level and statistical. The character layer is zero-width codepoints, Unicode tag characters, variation selectors, exotic spacing and typographic punctuation - removed by exact codepoint matching, verifiable by re-scanning, no model involved. The statistical layer is the vocabulary and cadence patterns, plus whatever the watermark contributes; that can only be addressed by rewriting.

Can invisible characters really be hidden in text?

Yes, and they survive ordinary copy-paste. Zero-width characters, Unicode tag characters and variation selectors all render as nothing while remaining fully present in the data. Run the free scan on any text and you will see the exact count.

Does the rewrite change my meaning?

It is instructed not to. Facts, figures, names, dates, quotations, citations, URLs, paragraph structure and claim strength are all held fixed, and any sentence without a marker is returned verbatim. Read the result before you use it — that is true of any tool that touches your writing.

Will this defeat a specific AI detector?

We make no claim about any named detector, and no claim to verified watermark removal. Detectors are proprietary, they change without notice, and they disagree with each other. Anthropic's watermark has no public detector at all yet. What we state precisely: the character-level markers are removed exactly and verifiably, and the wording markers in your scan report are what the rewrite targets.

Why is the scan free?

Because it costs us almost nothing. It is exact character matching, not a model call, so there is no reason to charge for it. You only spend words when you ask for an actual rewrite.

What happens to my text?

It is processed to produce your result and saved to your history so you can retrieve it later. You can delete any entry. It is never used to train a model.

How is usage counted?

By word, against the input text. Scans are free and never counted. Each plan also caps the words in a single request, so long documents need a higher tier or splitting up.

Can I use this on text I wrote myself?

Yes, and plenty of people do. Human writing picks up invisible characters from PDFs, CMS editors and word processors, and the typographic punctuation cleanup is useful on its own. The scan will tell you whether there is anything there.

AI watermarking law: what disclosure rules actually require

The regulations pushing providers to mark AI output, what machine-readable disclosure means in practice, and the gap between marking content and disclosing it.

Law and policy · 4 min read

Text watermarking went from research curiosity to shipped product across the industry in about a year. That was not a technical breakthrough. It was a compliance deadline.

What the rules require

The EU AI Act's Article 50 became enforceable on 2 August 2026. It requires providers of generative AI systems to mark synthetic output in a machine-readable format, so that it can be identified as artificially generated. Text is explicitly in scope, alongside audio, image and video.

Two words in there carry most of the weight.

Machine-readable rules out a visible label. A footer saying "generated with AI" is not what the regulation asks for, because it does not survive the content being copied elsewhere. The mark has to travel with the content itself.

Provider is the party the obligation falls on. That is the company running the model, not you.

Who is obligated

This distinction gets lost constantly, so it is worth stating plainly.

Providers — Anthropic, Google, OpenAI — must mark their output. That obligation is why Claude's text carries a watermark and why Gemini's carries SynthID.

Deployers, meaning businesses using these systems in their own products, have separate transparency duties: disclosing that users are interacting with an AI system, and labelling deepfakes and certain AI-generated public-interest content.

Individual users writing a document are not the subject of Article 50's marking requirement. The regulation does not impose a personal duty to preserve a provider's watermark in text you have been given.

That is a statement about this regulation specifically. It is not a general permission slip, and it says nothing about your university's rules, your publisher's contract or your employer's policy — which is usually where the real obligation sits.

Why it went global

Anthropic applied Claude's mark worldwide rather than to EU users only. Google did the same with SynthID.

Partly this is simpler than geofencing a model-level feature. Partly it is that other jurisdictions are moving in the same direction, and building two behaviours means maintaining two behaviours. Compliance tends to converge on the strictest applicable rule.

The practical effect is that a regulation with EU jurisdiction changed how text is generated everywhere.

Provenance standards

Two different technical approaches sit under "machine-readable disclosure", and they are not interchangeable.

Statistical watermarking for text. The mark is embedded in the content itself by biasing which words the model chooses. It survives copy-paste because it is the text. It degrades under paraphrase and translation, and it can only be read by whoever holds the key.

Signed metadata for images and other media — principally C2PA, the Content Credentials standard. A cryptographically signed manifest travels alongside the file recording how it was made. Anyone can verify it, because verification uses public keys. But it can be removed trivially: a screenshot, a re-encode, or most social platforms' upload pipeline will strip it.

The trade-off is symmetrical and unavoidable. Text watermarks are hard to remove and impossible for the public to check. Metadata is easy for the public to check and trivial to remove.

Anthropic uses both: statistical marking for text, C2PA for supported image formats.

What the rules do not require

They do not require detection to be public. Article 50 requires marking. It does not require providers to publish a mechanism, or to let third parties verify anything. Anthropic marks Claude's text and has not published detection details; that is compliant.

This produces the situation the whole field is now in: content is marked, and nobody except the marker can read the mark. Disclosure that only the discloser can verify is a strange kind of disclosure, and it is what the current rules produce.

They do not require the mark to be durable. No threshold is set for surviving editing. Providers document their own limitations, and every published system degrades under paraphrase.

They do not establish authorship. A mark records that a model was involved in producing some text. It cannot distinguish drafting from editing, and no provider claims otherwise.

Where this is heading

The gap worth watching is between marking and verification. A machine-readable mark that only its issuer can read satisfies the letter of a disclosure requirement while leaving publishers, educators and readers with nothing they can actually check.

Either detection becomes available to accredited third parties, or the requirement functions mainly as an internal capability for the providers. Anthropic has said technical documentation is forthcoming. What lands in it is the thing to watch.

Scan your text for the layer that is verifiable by anyone today: hidden characters, spacing and typographic punctuation, counted exactly.

AI watermarking law: what disclosure rules actually require

What the rules require

Who is obligated

Why it went global

Provenance standards

What the rules do not require

Where this is heading

Related

What C2PA is, and why it is not a text watermark

EU AI Act Article 50: why every provider is watermarking text now

How Claude's watermark works