Claude watermarks everything it writes.

Models launched on or after 2 August 2026 generate watermarked text. It is applied at the model level rather than in one app, so it comes with the model wherever the model runs. Anthropic says older models are in a transition period and it is working to add watermarking to those too. There is no opt-out published, no enterprise carve-out named, and no regional scoping: in Anthropic's words, "we're applying watermarking globally at launch because we don't yet have a durable way to scope it by region."

Files are a separate mechanism. When Claude produces a supported file type (a .png, .jpg or .svg), it attaches a content credential using the open C2PA standard. That is ordinary file metadata, and Anthropic is explicit that the credential says only that Claude was involved. It carries no identifying information. No account, no name, no prompt.

The driver is regulation, not a change of heart. Article 50 of the EU AI Act sets transparency requirements for AI-generated content, and the major model providers signed an EU Code of Practice on it in July 2026. A detection API is coming, which Anthropic says it is still working out.

Almost every post about this gets the mechanism wrong, and the wrong version is the one being sold. The watermark is not a zero-width space. Not a curly apostrophe. Not an em dash. Not a stray non-breaking space you can find and delete.

It is a version of SynthID-Text, the method Google DeepMind published in Nature in 2024. The idea underneath it goes back to a 2022 proposal by Scott Aaronson. Here is the whole thing in one sentence: when a model writes, there are usually several words that would work equally well in the next slot, and the watermark quietly biases which of those the model reaches for. Over enough text, that bias forms a pattern a detector can find.

Read that again with a marketer's eye, because it settles the question. The signal lives in the word choice itself. There is no character to strip, no metadata field to clear, no encoding trick to undo. It survives copy-paste because it survives inside the words you pasted. Any product claiming to remove it by cleaning characters is either confused or lying, and both versions take your money.

The characters are still worth cleaning, just for different reasons. Zero-width spaces and stray non-breaking spaces arrive with anything copied out of a chat window, and they quietly break search, word counts and diffs. Clean them because they are junk, not because they are the watermark.

This is the distinction that turns the panic down, and it is Anthropic's own framing: "a watermark can only determine that Claude was likely involved with the content at some point." Involved. Not wrote. Not ghost-authored. A hit on a page you researched, structured and rewrote says a model touched it somewhere, which is the same thing your browser history says.

Three limits do most of the work here:

• Short passages are unreliable. Detection needs volume. As Anthropic puts it, confidence rises with length, and it "doesn't work well on small samples". Your product descriptions and your captions are not the target.
• Factual writing carries almost no signal. The watermark is sparse where the word choices are constrained. Prices, limits, specs, numbers, code: there is no room to bias a $25 into anything other than a $25. The articles on this site are unusually dense in exactly that kind of sentence.
• Your own words carry nothing at all. Anthropic again: when "nearly all the words are the person's, there's very little (if anything) for the watermark to attach to."

So the honest read is that this is a provenance signal, not an accusation. The thing worth worrying about was never the mark. It was always publishing something no human had a real hand in.

Anthropic published the answer in one sentence, and it is the least convenient sentence in this whole story: "Light editing probably won't remove the watermark completely; a complete rewrite where every word is replaced will."

Sit with what that rules out. Running it through a paraphraser is light editing. Swapping ten adjectives is light editing. Asking a second model to "make this sound human" is light editing, and it also hands your text to a second model. The only thing on the list that works is the thing nobody selling a tool wants to say: writing it yourself.

Which is where this stops being a watermark problem. If your fix is a complete rewrite in your own words, you did not defeat a detector. You wrote the page. The detector became irrelevant on the way past, as a side effect of doing the work.

I write this site with Claude in the loop and I am not going to pretend otherwise. What changed on 2 August is nothing about the mechanics and everything about the margin for laziness. Here is the split that has always worked, now with a second reason to hold it:

1. The agent brings structure and research. You bring the sentences. Outline, the field to cover, the prices to go verify, the objection a reader will have. That is genuinely faster than doing it alone. Then the prose gets written, not accepted.

2. Every page carries facts only you have. This is the one that does the heavy lifting. When Metricool billed me $25 rather than the $18 I had published, no model on earth knew that. Same for the 10 charges on my card that all read APPLE.COM/BILL, and for the fact that the ClickUp I already pay for records video, which makes a $18-a-seat Loom the cheaper thing to cancel even though the swap is genuinely not the same. A page built on your own receipts cannot be generated, cannot be duplicated by a competitor, and has nothing generic left for a watermark to sit on.

3. Never ship a first draft as prose. Treat generated text as a structured outline that happens to be in sentences. The rewrite pass is where voice arrives anyway, and it is the exact operation Anthropic names as removing the signal. You get both for one pass.

4. Verify every number the day you publish. Not a watermark rule, just the rule. It has caught two wrong prices on this site that would have shipped otherwise.

None of that is a workaround. It is the same advice this site gave when the only tell was an em dash. The watermark did not create a standard, it just made the old one enforceable.

Not everyone has the same exposure here, and lumping them together is how the panic spread.

Your own site, your own newsletter, your own product copy. Low stakes. You are the publisher, nobody is auditing you, and the practical worry was never a watermark. It was that thin generated pages do not rank and do not sell.

Anything submitted to someone who checks. Coursework, client deliverables where AI use is contractually restricted, applications that ask you to declare it. This is the real category, and the answer is not a stripping tool. It is doing what you agreed to do, and disclosing when asked.

Anything where a false accusation would cost you. Worth knowing the failure runs both ways: the mark proves involvement, not authorship, and heavy human editing degrades it. Someone waving a detector result at your work is holding weaker evidence than they think, and now you know exactly how weak.

What nobody should do is buy a remover. There is no character to remove, and anyone selling one is describing a mechanism that does not exist.