Phravv

Paste text. Get a high-fidelity paraphrase.

Same meaning. New wording.

What Claude watermark claims can and cannot do

Claude can leave a statistical pattern in the words it chooses. Anthropic says a light edit probably will not clear that pattern, and replacing the words can weaken it. That is not a guarantee. Phravv does not detect the pattern, does not promise it is gone, and never rewrites with Claude. This page is not a detector-bypass guide.

. Jorge Bustamante, Edmonton.

What the watermark is

When Claude writes, it can leave a statistical pattern in the words it chooses. Anthropic’s method is a version of Google DeepMind’s SynthID-Text, from a family of ideas that goes back to a 2022 proposal by Scott Aaronson. The pattern lives in low-stakes word choices. Nothing is inserted into the file. There are no hidden characters. You cannot see it on the page. Anthropic described it on August 14, 2026.

They said they were doing it to comply with EU marking rules, applied globally because they could not yet limit it by region. The mark does not add tokens, and a reader is not meant to notice it. A style detector looks at phrasing and has no key. Phravv does not have the key, and it does not guess.

Anthropic: How Claude’s text watermark works.

What a positive result would mean

A check can only speak to likelihood. Anthropic’s question is whether Claude was likely involved at some point. A found mark does not name a person, a company, or a chat. It does not separate “Claude wrote this” from “Claude heavily edited this.” It does not say the text was human, and it does not identify a different model.

A miss is not proof either. Short passages give you less to go on. Models launched before August 2, 2026, are on a transition Anthropic describes, with the current list on their help page.

Anthropic: How Claude’s text watermark works; How Claude marks AI-generated content.

Where the pattern has less room

The pattern needs choices. Where only one next word is acceptable, there is nothing to bias. A version number, a function name, a quote, or the only right date is that kind of spot. Ordinary prose, where many phrasings would do, gives the method more room.

Anthropic makes the same point about code: exact code is generally thinner than prose, while a comment can be phrased more than one way. A light proofread is thin for the same reason. A translation Claude produces is marked, they say, because Claude chose every word. Phravv is not a translator.

Whether a date or a quote survived a paraphrase is a different question from whether a statistical pattern moved. The fixture text for the paraphrase checks is on the related page.

Anthropic: How Claude’s text watermark works.

Related: What Phravv preserves

What Anthropic says a rewrite can do

Anthropic’s own answer is plain. Light editing probably will not remove the watermark completely. A complete rewrite, where the words are replaced, will. They add that once every word has been replaced, it is fair to ask whether the text is still the generated text.

That is their claim about their mark. It is not a guarantee from me, and it is not a result Phravv checks. I do not know whether a given paste still carries a pattern, before or after a rewrite. I will not publish settings, tricks, or success rates for clearing one. This is not a detector-bypass guide. The pair below only shows how much wording can move. I wrote both versions. Neither is a detection result.

Before

The support queue closed at 42 tickets on June 3, which is the lowest count this quarter.
A one-word edit

After

The support queue closed at 42 tickets on June 3, which is the smallest count this quarter.

42 and June 3 stay. Anthropic says a light edit like this probably will not clear a mark. I am not confirming that here.

The same facts, with the sentence rebuilt

After

On June 3 the support queue finished at 42 tickets, the lowest count this quarter.

The count, the date, and the quarter claim stay. Anthropic says replacing the words can remove the pattern. That is not a promise about your draft.

Anthropic: How Claude’s text watermark works.

What Phravv does not do

Phravv does not detect the mark. There is no badge and no score. Anthropic’s detection API is in private preview for eligible organizations, as of the September 1, 2026 update on their news post. It is not a public checker, and Phravv does not call it. The rewriter will not depend on a detector.

Phravv never rewrites with Claude. If Claude rewrites the text, it can put the same kind of mark back in. The model behind the rewriter is not an Anthropic model. Mixing untouched sentences back into a rewrite can keep a signal a full pass would have moved. That is not a procedure.

Anthropic: How Claude’s text watermark works.

Text marks and file marks are different

Claude can also attach a content credential to some files, such as a PNG or a JPEG. That is C2PA metadata in the file, not the statistical pattern in prose. Phravv never sees it. You paste text. Changing a paragraph does not edit an image’s metadata. If your question is about a file, Anthropic’s help article is the right page.

Anthropic: How Claude marks AI-generated content.

How to read this

Use the rewriter because you want a second pass that keeps the meaning. Read Anthropic’s notes if you care about the mark. The claims on this page come from their news post and their help article.

Treat any “guaranteed removal” claim, mine or anyone else’s, as marketing, not evidence. I do not have a guarantee to sell. The useful question here is still the editorial one. Did the meaning stay, and did the facts stay.

Anthropic: How Claude’s text watermark works; How Claude marks AI-generated content.

Rewrite a draft