Skip to content

Anthropic

Paraphrasing won’t erase a watermark, declaude’s tests say

8 min read

Written by an AI model from the linked sources and checked against them by a second model. How we work

On August 16, James Padolsey of NOPE published an interactive guide, written for Declaude, a rewrite tool. It explains where a watermark can hide in plain text and measures how much rewriting it takes to remove one.

This now applies to anyone who uses Claude daily. As of August 2026, new Claude models mark their text at the model level, and earlier models will follow. Anthropic describes its version in an announcement and a help page.

Where does a mark go in plain text?

It doesn’t go in the characters. Anthropic says nothing is added to the text and there are no hidden characters. The mark sits in the choices between words. Partway through a sentence, a model doesn’t have a single “next word”. It has a weighted shortlist of candidates that would each work, and it samples from that list. Padolsey’s example is a sentence one word from the end, “The results of the study were quite”, where several endings are equally fine. A page of text holds hundreds of these forks, one per word, and at many of them the model has real room to choose. Whoever controls how the pick lands can write a pattern into the text without changing what it says.

The classic recipe comes from Kirchenbauer et al. (ICML 2023). At each fork, maths seeded by a secret key splits the candidates into “green” and “red”, and only the key holder can reproduce that split. Sampling is then tilted slightly toward green. A red word can still win, just a little less often. A word’s colour isn’t fixed, either. The key computes it from a short run of the preceding words, so the same candidate can be green after one prefix and red after another. What accumulates across the text is only a lean toward green, and only the key holder knows which words were green at which position.

Google’s SynthID-Text, the scheme Google runs in production (Dathathri et al., Nature 2024), reaches the same result by another route. A few candidates are drawn from the model’s own odds, the key scores them in a small tournament, and the bracket is arranged so that, averaged over the key’s draws, every word’s odds stay exactly what the model intended. Aaronson’s scheme, built at OpenAI, derives the random draws themselves from the key. The maths differs, but in Padolsey’s words, “the mark lives in the choices.”

Anthropic’s description of Claude’s mark fits the same family. Where a normal sampler would use an arbitrary random number generator, the key plus a few preceding words decides which word Claude picks. Anthropic’s example is a low-stakes choice such as “overcast” versus “grey”, and it says the watermark doesn’t push Claude toward a word it wouldn’t have considered anyway. The company also says the mark has no practical impact on output quality, readers can’t tell marked from unmarked text, and it needs no extra tokens, so it costs nothing more. The exact scheme hasn’t been disclosed.

Counting greens

Detection doesn’t read the text or judge its style. The detector replays the key’s colouring over the words and counts how many came up green. In unmarked text, or with the wrong key, green comes up about half the time, as with a coin flip. Marked text comes up green too often for chance to explain it. Anthropic describes its check the same way: with the key, you test whether a sequence of words is consistent with the choices Claude would make, and you get a probability that Claude generated it.

Padolsey’s demo draws the tilt strong enough to see. A production mark leans far more gently and needs correspondingly more text. Even in the demo’s 50/50 model, a 1,500-word document would be flagged at only about 55% green. A small lean becomes persuasive only over length, which is why short texts are hard to call. Text with a single right continuation also gives the mark too little room to work with, and the guide names code, quotations and lists of facts.

What editing does to it

Each word’s colour comes from the one to a handful of words before it, depending on the scheme. A position only counts as evidence if a short window of the original wording, the word plus its neighbours, survives intact. Editing erases the mark where those runs break and nowhere else. The guide’s slider shows how much of the text a heavy edit still leaves standing, and how far toward a full rewrite you have to go before the evidence actually dies.

The published research describes the same pattern. Light or one-pass paraphrase dilutes the mark without deleting it. In Kirchenbauer et al.’s follow-up experiments (ICLR 2024), the detector recovered given enough text, and even human paraphrase became detectable again after roughly 800 tokens, about 600 words. Anthropic’s help page is more guarded: it says the mark travels with copied and pasted text and “may persist through some editing”.

Declaude ran its own tests with a known key, using MarkLLM’s KGW and EXP schemes on an open model. After the tool’s full-rewrite route, which recomposes text from its meaning, about 0.5% of windows survived. Detection AUC fell from 0.99 to about 0.5. AUC is a score where 1.0 is a perfect detector and 0.5 is guessing. The guide models leftover evidence as z ≈ f·√N·z₁, where f is the surviving fraction, N is document length and z₁ is per-token strength. Losing most of f is therefore what kills detection.

Not every mark collapses that way. Schemes keyed on the word itself, with no context, held up much better in the same tests: AUC stayed between 0.73 and 0.84 after re-composition, because a same-meaning rewrite keeps enough of the same words. Their weakness is different. A colouring reused everywhere can be reverse-engineered from enough output. Other schemes hide the mark in the meaning, which a same-meaning rewrite partly keeps, and the only answer the guide knows for those is regenerating the text from an outline.

The guide says outright that all of these figures come from open implementations. With Anthropic’s scheme undisclosed, no one outside Anthropic can yet run the test against Claude’s own mark. Open-source tooling is waiting on the same thing: WMTrace, a forensic workbench, won’t ship an Anthropic detector until there is an official verification API or a reproducible public spec.

What changes for Claude users

The mark is applied at the model level, so it is present whichever Claude product or surface produced the text. Anthropic says it is watermarking to comply with the EU AI Act. As of August 2, AI providers serving the EU market must mark AI-generated content, and Anthropic says other major model developers have signed the same Code of Practice. Its help page lists text watermarks for several models released before that date, including Opus 4.8, Sonnet 5 and Fable 5. It also says the rollout on Amazon Bedrock will complete by October 17. Supported files such as PNG or JPEG get Content Credentials (C2PA), signed provenance metadata, instead of a statistical mark.

Google has watermarked text from the Gemini app and web experience since 2024. At the time the guide was written, its API was a documented exception.

Only the key holder can run the check. A teacher, an editor or a public “AI detector” site can’t. A genuine check needs the provider’s key or a checking service the provider runs. Google runs an early-access SynthID detector portal. Anthropic’s detection is still only in private preview, open to eligible organizations as EU law requires: regulators, law enforcement, media, fact-checkers, independent researchers, educational organizations and EU civil society groups. Padolsey warns against confusing this with style-based tools such as GPTZero, which he calls famously unreliable. A watermark check is a deliberate statistical test that only works with the key.

A clean result proves little. Text from older models, or heavily edited text, comes back clean even when an AI wrote it.

A found mark, in Padolsey’s summary, means “processed by”, not “written by”. Anthropic’s own documentation says that human text Claude merely proofread or translated picks up the mark, and the help page words a detection as content that “may have been generated or processed by Claude”.

No comments yet. Be the first.

Join the conversation

Sign in to leave a comment. It appears after moderation.