How to beat AI detectors and the new ChatGPT and Claude watermarks

ChatGPT and Claude now hide a watermark inside the words they write. Copying the text into Notepad doesn't remove it. Neither does deleting em dashes or running it through a "humanize" prompt in the same AI.

What works is rewriting enough of the words yourself. This guide covers how the watermarks and AI detectors catch you, what doesn't work, and the steps that do.

The quick answer

The three ways AI writing gets caught

There are three separate checks, and each one needs a different fix.

Most people only think about the second one. The first one is new, and it's the one copy and paste can't touch.

The new watermarks: how ChatGPT and Claude mark your text

How a word watermark works

Every time an AI writes, it picks the next word from a handful of good options. "The results were great", "the results were strong" and "the results were solid" all work.

Normally that pick is random. With a watermark, a secret key makes the pick instead. One word tells you nothing. Over a few hundred words, the AI keeps leaning towards the same "secret key" choices, and the company's detector can spot that pattern.

The words read completely normally. There's no hidden symbol and nothing to delete. The mark is the words themselves, which is why it survives:

• Copying and pasting into Notepad, Google Docs, Word or an email
• Changing the font, formatting or layout
• Deleting em dashes or other punctuation
• Fixing a few words here and there

The watermark doesn't contain your name, account or chats. It only says "this AI was involved".

Claude's watermark

Anthropic announced it on 11 August 2026 to meet new EU AI rules.

• Where: Worldwide, in every Claude product: the Claude app, Claude Code, Cowork and the API.
• Which models: Claude models launched on or after 2 August 2026. Anthropic is adding older models over the coming months.
• How it works: A secret key shapes the word choices, as described above.
• Proofreading: Light proofreading of your own writing usually leaves too little to detect. The more Claude rewrites, the stronger the mark.
• Translation: Translations are watermarked, because Claude picks every word.
• Files: Images and some files Claude makes get a "made with AI" label in their file data (called C2PA) instead.
• What removes it: Anthropic says light editing probably won't. A full rewrite that replaces every word will.

ChatGPT's new watermark (textGrain)

OpenAI announced its watermark, called textGrain, on 5 October 2026.

• Where: ChatGPT and Codex for people in the EU, on every plan. It's rolling out over a few weeks.
• Outside the EU: It isn't on by default in ChatGPT yet. Developers using OpenAI's API anywhere can switch it on, but it starts off.
• How it works: The same idea as Claude's. A secret key shapes the word choices.

OpenAI published how well it holds up. In its tests on answers of about 300 words:

OpenAI also says short passages, maths answers and translated text are harder to detect. So a quick caption is far safer than a 1,000-word essay.

Who can check for a watermark

Right now, nobody can paste your text into a free website and check for a ChatGPT or Claude watermark.

• Claude: Anthropic's checker is in private preview. Regulators, law enforcement, media, fact-checkers, researchers, education organisations and some businesses can apply.
• ChatGPT: Only researchers and expert organisations that OpenAI approves.
• Turnitin and GPTZero: They don't have the secret keys, so they can't read these watermarks. They use their own style checks instead.

Two things to keep in mind. Education organisations are on Anthropic's list, and access will widen over time. And a watermark doesn't fade. Text you hand in today can be checked next year. If your school, uni or employer bans AI, none of these steps changes the rule.

How AI detectors read your writing

Detectors like Turnitin, GPTZero, Pangram, Originality.ai and Copyleaks don't look for watermarks. They're AI models trained on millions of examples of human and AI writing. They look for the habits AI can't help.

That last row matters. Turnitin added "AI bypasser" detection in August 2025 to catch text from humanizer tools, and Pangram has trained on humanizer output too. Detectors now learn the fingerprints of the tools that try to beat them.

You get a percentage back. That number is the detector's best guess, not proof.

Detectors get it wrong, a lot

Detector companies say they wrongly flag human writing less than 1% of the time. Independent tests disagree:

• A Stanford study found AI detectors flagged more than half of essays written by people who speak English as a second language as AI-written.
• Plain, simple, very correct writing scores as "more AI" because it's predictable. Technical writing, legal writing and lists are often flagged.
• Results swing between detectors. The same essay can get very different scores from two tools.
• Curtin University in Perth turned off Turnitin's AI detection from 1 January 2026.

So a high score doesn't mean you used AI, and a low score doesn't mean a watermark is gone. They're separate checks.

What doesn't work

The Humanizer skill I cover in the signs of AI writing guide is great for making AI text nicer to read. Its own creator says beating detectors isn't the goal, and detectors still flag most of its output.

What works: step by step

The only thing that beats all three checks is changing enough of the actual words. The best way to do that is to make the ideas and wording yours from the start.

Step 1: Put your own thinking in first

The more of the text that starts with you, the less there is to catch. Before you ask AI to write anything, give it your notes, your opinion and your examples. Talking is faster than typing, so use voice mode if you like.

This prompt gets AI to pull your ideas out of you, in your words:

I need to write [WHAT YOU'RE WRITING, E.G. A BLOG POST ABOUT OUR NEW SERVICE]. Don't write it for me. Interview me instead. Ask me one question at a time about what I want to say, my opinions, real examples and anything only I would know. When I say "done", give me back my answers as rough bullet points using my exact words.

Step 2: Use AI to shape your words, and keep it light

If AI only tidies your writing, there's very little of its own word choice in the result. Ask it to keep your words:

Here's my rough draft: [PASTE]. Fix spelling, grammar and the order of the points. Keep my exact words and phrases wherever you can. Don't add new sentences, don't swap my words for fancier ones, and tell me what you changed.

If you've had AI write a full draft already, that's fine. Steps 3 and 4 matter more for you.

Step 3: Rewrite it in your own words

This is the step that breaks the watermark. The 25% figure from OpenAI's tests shows that the more words you change, the weaker the mark gets. Aim to change most of them.

The method that works best is read, close, rewrite:

1. Read one paragraph of the AI draft.
2. Close it or scroll it out of view.
3. Write that point from memory, the way you'd say it out loud to a friend.
4. Move on to the next paragraph.

You'll naturally use your own words, change the sentence shapes and drop the filler. It's slower than editing on top of the draft, but editing on top keeps most of the AI's words.

To find what still needs work, ask AI to point out problems without fixing them:

Read this and list every sentence that sounds AI-written. For each one, tell me why (word choice, rhythm, structure or filler) in a few words. Don't rewrite anything. I'll fix them myself.

[PASTE YOUR TEXT]

Step 4: Break the AI habits

Go through your rewrite and fix the habits detectors look for:

• Mix your sentence lengths. Some short. Some long, with a bit of a ramble in them, the way people actually talk.
• Cut the AI words. Crucial, seamless, robust, leverage, delve. The full list is in the signs of AI writing guide.
• Drop the neat ending. Stop when you've made your point. No "in conclusion".
• Change how paragraphs start. Start some mid-thought, some with a question, some with a one-liner.
• Use the words you use. If you'd say "heaps" or "pretty good", say it.

Step 5: Add what only you know

This does more than anything else, for detectors and for readers. AI can't fake specifics it doesn't have.

• A real number: "we tested 14 hooks" beats "we tested many hooks"
• A real story: a client, a mistake, something that happened last week
• A clear opinion, even one people might disagree with
• Names, places and dates

Step 6: Clean hidden characters and file data

Some ChatGPT models were caught in 2025 putting invisible special spaces into longer answers. They look like normal spaces but can show up in some checkers. They aren't the watermark, but they're worth removing.

If you use Claude Code, paste this in with your file:

Check [FILE NAME] for invisible or unusual characters: zero-width spaces, narrow no-break spaces, non-breaking spaces, soft hyphens and letters that look normal but come from another alphabet. Replace each one with its normal version using a script, without rewording anything. Show me what you found and where.

For images and documents, the "made with AI" label sits in the file data. Exporting a screenshot of an image drops most of it. The Remove AI Marks skill below strips it properly.

Step 7: Test it, then keep your proof

Run your final version through two free detectors, like GPTZero and Originality.ai's free checker. If one flags a paragraph, rework that paragraph. Don't chase a perfect 0%. Plenty of real human writing gets a small AI score.

Then keep your proof. Write in Google Docs or Word with version history on. Docs saves every change automatically, so you can show exactly how your writing came together (File > Version history > See version history).

The Remove AI Marks skill

Remove AI Marks is a free, open-source skill for Claude Code (skills are instruction files that teach Claude a specific job). It has more than 25,000 stars on GitHub. It works in two layers.

The rewrite layer follows the same rule as step 3: enough of the words have to change. Set it up with a different AI from the one that wrote your text, or the rewrite can add a fresh watermark. The creator recommends a free AI that runs on your own computer through Ollama, so your text never leaves your machine.

How to install it

It runs a small helper program on your computer, so it takes a few minutes to set up. Claude Code can do the whole thing for you.

1. Open Claude Code in your terminal (the text window where Claude Code runs).
2. Paste these one at a time:

/plugin marketplace add guillaumemeyer/watermarks-remover

/plugin install watermarks-remover@watermarks-remover

3. Then paste this to set up the helper and the rewrite AI:

Set up the watermarks-remover project from https://github.com/guillaumemeyer/watermarks-remover so I can use the remove-ai-marks skill. Clone it, start its local service on 127.0.0.1:8765 using the README, and check it's running with the health check. For the rewrite step, install Ollama with a small free model and configure it as the rewrite backend. Explain each step in simple terms and stop to ask me before installing anything.

4. Once it's running, use it like this:

/remove-ai-marks Clean this text and rewrite it to break up any AI watermark. Keep my meaning and my tone. Then tell me what you removed and what you changed.

[PASTE YOUR TEXT OR GIVE THE FILE NAME]

If you only want the clean-up layer, the same plugin includes a lighter skill called clean-user-facing-text. It strips hidden characters without the helper program or a rewrite AI.

Treat the rewrite as a first pass. Then do steps 3 to 5 yourself. Your own words and specifics are what make it hold up, and they make it better to read.

What about images?

Images carry their own marks, separate from text.

Instagram and Facebook read the file data and can add an "AI info" label to your post.

Does Google penalise AI content on your website?

Not for being AI. Google's guidance says it rewards helpful content however it's made. What it does penalise is pumping out lots of pages mainly to rank, with little value for readers. Google calls that "scaled content abuse".

For a business website, the steps above help anyway. Your own examples, opinions and specifics are what make a page worth ranking.

If you've been wrongly flagged

Detectors get it wrong often enough that you should know what to do.

1. Stay calm and ask for the details. Which tool flagged it, what score, and which parts.
2. Show your version history. Google Docs and Word show every change you made over time, which is strong proof you wrote it.
3. Bring your notes and drafts. Outlines, research links and earlier versions all help.
4. Point out that a score isn't proof. Turnitin itself says its AI score shouldn't be the only basis for action against a student.
5. Offer to talk it through. Explaining your work in person is the best proof there is.

FAQ

Quick recap

1. ChatGPT (in the EU) and Claude (worldwide) now hide a watermark in their word choices
2. Copy and paste, deleting dashes and "make it sound human" prompts don't remove it
3. Detectors like Turnitin check your style instead, and they wrongly flag real writing often
4. Put your own thinking in first, and keep AI edits light
5. Rewrite the draft from memory, paragraph by paragraph
6. Break the AI habits and add specifics only you know
7. Clean hidden characters and file data, or use the Remove AI Marks skill
8. Test with two detectors and keep your version history