AI Guides › Playbooks
By Nigel Guy · 7 min read
Since Claude started watermarking its text, the most common reaction has been to look for a trick that removes the mark: swap some synonyms, ask another chatbot to "humanise" it, run it through a paraphraser. That feels like a fix because the words change. But it treats the problem as a detection problem when it is really an authorship problem, and the most popular trick, asking an AI to rewrite it, may simply put a fresh mark on top.
The rule: don't try to scrub the mark. Decide how much of the piece should be yours, do that much of the writing yourself, and say plainly what Claude did.
| Question | Status at time of writing | Source |
|---|---|---|
| Is the watermark real? | Yes. Supported Claude models embed an imperceptible watermark in the text they generate. | Claude Help Centre |
| Since when? | Models launched on or after 2 August 2026 support marking from launch. Older models are being added over the coming months. | Claude Help Centre, Anthropic |
| Does it apply in the UK? | Yes. Anthropic applies it worldwide and on every surface: Claude apps, Claude Code, Cowork, the API and the cloud partners. | Claude Help Centre |
| Hidden characters? | No. Nothing is added to the text. | Anthropic |
| Can it identify you? | Anthropic says no. The mark contains no user, organisation or chat information. | Anthropic |
| Can you check your own text? | Not for text. Watermark detection is a private-preview API for eligible organisations such as regulators, media, researchers and educators. The free Claude Content Checker reads file credentials (C2PA), not text watermarks. | Claude Help Centre |
| What exactly removes it? | Not specified. Anthropic says light editing "probably won't" remove it and a complete rewrite in which every word is replaced will. It publishes no figure for anything between those two. | Anthropic |
Model coverage varies. The help centre's table currently shows text watermarks for Fable 5 and 5.1, Mythos 5.1, Opus 4.8, 5 and 5.5, and Sonnet 5 and 5.5. Opus 4.5 to 4.7, Sonnet 4.5 and 4.6, Haiku 4.5 and Mythos 5 show file credentials only for now. The note says some Amazon Bedrock rollouts complete by 12 October.
When Claude writes, it picks the next word from a shortlist of sensible options. Often two options are equally good, for example "grey" or "overcast", and a random number decides between them. With watermarking, that randomness comes from a secret key combined with the few words just before. The text reads the same. Across hundreds of these small choices, though, a pattern builds up that someone holding the key can test for. Anthropic says the method is a version of Google DeepMind's SynthID-Text.
What follows from that:
This is a four-rung ladder. Before anything leaves your hands under your name, pick the rung it sits on and apply that rung's step.
| Rung | What Claude did | Likely mark | What you do |
|---|---|---|---|
| 1. Checked | Spotted typos and grammar in your draft | Little or none | Accept the fixes. No pass needed. |
| 2. Shaped | Suggested structure, cuts and questions, but you wrote the sentences | Little | Note the help if your context expects it. |
| 3. Drafted, then rebuilt | Wrote a first draft that you then rewrote in your own words | Falls the more you rewrite. Unknown in the middle. | Run the pass below in full. |
| 4. Claude's words | You are sending Claude's text largely as written | Present | Don't pretend otherwise. Disclose it where honesty or the rules require. |
The pass itself, for rung 3:
A worked example, entirely hypothetical: you run a two-person bookkeeping firm and ask Claude for a 600-word client newsletter piece on Making Tax Digital deadlines. The draft is fine but generic. You list its five points, close it and write 450 words from the list. You add the reminder you give every client about digital record-keeping and a line about your own Friday cut-off. Two sentences still read like Claude when you say them aloud, so you redo them. The result sits on rung 3. You add "drafted with AI help, checked and rewritten by us" because your engagement letter promises transparency. Whether a detector would still pick up a trace is unknown, and with this approach it doesn't matter.
The tempting shortcut is "rewrite this in my voice". Don't use it on this rung. A rewrite by any supported Claude model is new Claude text, and Anthropic says other providers that signed the same EU code are adding their own watermarks. Use Claude to point out where your rewrite is unfinished, and do the writing yourself.
You are an editing coach helping me make a piece of writing genuinely my own. You must not write replacement sentences for me.
Context: I used an AI draft as a starting point and have now rewritten it. My aim is that the final wording, examples and rhythm are mine, so that I can honestly put my name to it.
Inputs:
- AI draft: [PASTE_AI_DRAFT]
- My rewrite: [PASTE_YOUR_REWRITE]
- Who will read it and where: [AUDIENCE_AND_CHANNEL]
- Three samples of my normal writing: [VOICE_SAMPLES]
Steps:
1. Compare my rewrite with the AI draft. List every sentence in my rewrite that keeps the draft's structure or most of its wording, quoting only the first few words of each.
2. For each one, ask me a single question that would help me rewrite it from my own knowledge, such as "What happened the last time a client missed this?". Do not suggest wording.
3. Name up to three spots where my rewrite sounds unlike my voice samples, and say why in one line each.
4. Name any claim in my rewrite that I should check before sending.
Output: three short numbered lists headed "Still the draft's wording", "Doesn't sound like you" and "Check before sending". Nothing else.
If any input is missing, ask me for it before you start. Before you answer, check that you have not written a single sentence I could paste in.
Fill in the AI draft, your rewrite, the audience and three short samples of your own writing, such as emails or posts.