Claude's Text Watermark: How It Works & What It Means (2026)

What is Claude's text watermark?

A text watermark is a hidden statistical pattern woven into generated text that a reader cannot see but that a party holding a secret key can detect. For Claude, it provides a way to later ask, with some probability, whether a given passage was written by Claude. It is not unique to Anthropic: as of early August 2026 the EU requires AI providers serving its market to mark AI-generated content, and other major model developers who signed the same commitment are adding their own watermarks too.

How does the watermark work?

Language models generate text one word at a time, and at each step they choose from a list of plausible next words. Often several options are equally reasonable, and the pick between them is settled essentially at random. Take a sentence ending in the weather was cold and: overcast and grey are both fine, and which one appears normally comes down to a random number.

Watermarking works by changing the source of that randomness. Instead of an arbitrary random number, the model uses a secret key together with the preceding words to decide among the equally good options. The choices still look random, but they now follow a hidden pattern that anyone with the key can check for. Importantly, the watermark only nudges between words the model would already consider; it does not push Claude toward strange or wrong words it would never normally use, and it does not override cases where one specific word is clearly correct.

Does the watermark affect Claude's output?

No, not in any way a reader would notice. A watermarked response is indistinguishable from an unwatermarked one, and nothing visible is added, unlike the watermarks on banknotes or documents. There are no hidden characters inserted into the text either. The method is a version of the SynthID-Text approach that Google DeepMind published in Nature in 2024, part of a family of techniques dating back to a 2022 proposal, all built on the same idea of only changing the source of randomness. In testing, both large-scale user ratings and side-by-side human comparisons found no meaningful difference in quality, creativity, or readability between watermarked and unwatermarked text.

Where the watermark is weak: code, facts, and edits

The watermark can only live in genuinely free choices between equally good words, so it thins out or disappears wherever an exact output is required. A few cases behave differently:

• Facts and math: where only one answer is correct, such as completing 2 + 2 or naming a specific work, there is no free choice for the watermark to use.
• Code: code usually has to be exact, so it carries far less watermarking. The pattern can only attach to arbitrary choices like the wording of comments, which has a negligible effect on the code itself.
• Proofreading and light edits: when Claude only lightly edits text a person wrote, almost all the words are the person's, so there is little or nothing for the watermark to attach to.
• Translations: these do carry a watermark, because Claude chooses every word.

A general rule follows from this: the more Claude writes, the more choices it makes, and the stronger and more detectable the watermark becomes. On short passages, confidence is low.

What it means for users

For day-to-day use, the practical impact is close to zero. Watermarking does not slow the model down and produces no extra tokens, so it costs the same to run and use. It also carries no identifying information: there is nothing in the watermark or its key that could reveal anything about you, your organization, or your conversations with Claude. And it does not change ownership or legal responsibility for an output; it only helps test whether Claude was involved, not who owns the result.

Why Anthropic is adding a watermark: the EU AI Act

The reason is regulatory. Anthropic, alongside around 190 signatories including several other major AI providers, signed the EU Code of Practice on Transparency of AI-Generated Content in July 2026, which requires providers to mark AI-generated text. Anthropic is applying the watermark globally at launch because there is not yet a durable way to limit it to a specific region. Models launched before August 2, 2026 fall under a transition period, and watermarking will be added to those older models over the coming months.

Files, and a coming detection tool

Text is only part of the picture. When Claude produces a supported file type such as a PNG, JPG, or SVG, it attaches a content credential, a small cryptographically signed note in the file's metadata stating the file was made or processed with Claude. This uses the open C2PA industry standard, the same one cameras and photo editors use to record an image's origin, and any C2PA-aware tool can read it. This is not a watermark: nothing inside the file changes, and like the text watermark it carries no identifying information.

For checking text, Anthropic says a watermark detection API is coming, with details still being finalized. This differs from third-party AI detectors, which do not have the key and instead guess from stylistic tells in the phrasing rather than reading an actual watermark.

Limitations and whether it can be removed

The watermark is a probabilistic signal, not proof of authorship. It can only indicate that Claude was likely involved at some point, and it cannot distinguish writing that Claude produced from writing it merely edited heavily. It does not confirm that text is human-written, and it cannot identify a different AI, which would use a different key or method entirely. As for removal, light editing probably will not strip the watermark, but a complete rewrite where every word is replaced will, at which point it is debatable whether the result is still AI-generated at all.