ChatGPT Text Watermark EU: What Changed and What to Do

OpenAI’s ChatGPT text watermark is rolling out in the EU, on all plans, for both ChatGPT and Codex. If you write in English and edit in the EU, your output may now carry a statistical mark in its word choices. It is not a hidden character you can delete. It is a pattern in how the model picks words. Rewriting is still the fix.

What actually changed this week?

OpenAI announced that ChatGPT and Codex output in the EU will carry a watermark over the coming weeks, on all plans. The company says it is “not making t” — the snippet cuts off there, but the practical point is that this is a platform-level change, not a setting you can toggle off. If you are in the EU and you used ChatGPT this week, your drafts are affected.

The watermark lives in the model’s word choices, not in hidden characters. That is the detail most people miss. You cannot strip it with a find-and-replace for zero-width spaces. You cannot see it in a hex editor. It is embedded in the statistical fingerprint of the text itself.

According to writehuman.ai’s report on the change (https://writehuman.ai/blog/chatgpt-text-watermark-eu), OpenAI’s own tests show that swapping a quarter of the words for synonyms cut detection from about 92% to 17%. That is the single most useful number in the whole announcement. It tells you the watermark is fragile in exactly the way human editing is strong.

Who is actually affected?

If you write for a living in the EU, this is your problem now. That includes:

• Students submitting essays, theses or take-home assignments written with ChatGPT assistance.
• Marketers and content teams publishing ChatGPT-assisted copy on EU-facing sites.
• Non-native English writers who use ChatGPT to smooth grammar and syntax, then submit the output.
• Researchers and grant writers drafting sections with model help.
• Freelancers delivering ChatGPT-assisted work to EU clients.

The watermark does not care whether the ideas are yours. It tracks the model’s word choices. If you prompted ChatGPT, pasted the output, and changed nothing, the mark is there.

The uncomfortable part: detection is not proof of authorship. A watermark is a signal, not a verdict. OpenAI’s own detection rates vary by task. In the writehuman.ai report, 400-token answers to psychology questions were detected about 95% of the time, 200-token answers about 80%, and math answers substantially lower. Short text is harder to mark and harder to detect. A watermark does not tell a reader who wrote your paragraph. It tells them a model was involved somewhere in the chain.

If you are staring at a flagged draft right now, start with our guide to reading the score before you rewrite anything: AI Detector Flagged Your Writing. It walks through what a sentence-level score actually measures, so you fix the lines that carry the signal instead of rewriting everything.

What the watermark does and does not do

A watermark is not a plagiarism flag. It is not a claim that you cheated. It is a statistical pattern that a detector can look for. The confusion between those three things causes most of the panic.

Here is what the watermark changes in practice:

What people assume	What the watermark actually does
Hidden characters you can delete	Lives in word choices, not characters
Proves you used ChatGPT	Signals model involvement, not authorship
Survives any editing	Weakens when you rewrite in your own words
Applies to all languages equally	Strongest in some languages, weakest in others
Blocks your work from being published	Changes how detectors read your text

The writehuman.ai report notes that detection rates vary by language, with Romanian the weakest of the same 24 languages tested. If you write in English, you are in the stronger-detection group. That cuts both ways: the mark is more likely to be present, and more likely to be found.

What to do today

You do not need to panic. You need a routine. Here is the one that works for a draft you already wrote with ChatGPT.

1. Save the original. Before you touch anything, save the ChatGPT output as a separate file. You may need to show your process later.
2. Read it out loud. Every sentence that sounds like a press release gets marked. If you would not say it to a colleague, rewrite it.
3. Change the words that are not yours. Not synonyms for their own sake. Replace abstract phrasing with the specific thing you mean. “Leverage synergies” becomes “use the same supplier.”
4. Add what only you know. A date, a name, a number from your own data, a caveat you learned the hard way. The watermark cannot mark facts it never saw.
5. Check your claims. Every quantity, citation, negation and qualifier. Rewriting is not worth it if you change what the sentence asserts.
6. Re-score at sentence level. See which lines still carry the signal. Fix those. Leave the rest alone.

Step three is where the 92% to 17% figure matters. Swapping a quarter of the words for synonyms cut detection in OpenAI’s tests. You are not trying to hit a quarter. You are trying to make the text yours. The number just tells you the mark is not permanent.

When this is the wrong approach

If you are writing an academic essay and your institution prohibits AI assistance, editing the output does not make it allowed. The watermark is a technical signal. The policy is a rule. Rewriting a prohibited draft does not resolve the policy question.

If you are in a dispute about authorship, do not rewrite the evidence. Preserve the original, document your process, and follow your institution’s or employer’s procedure. Our guide on academic integrity hearings explains what panels actually assess.

If you are writing in a language where the watermark is weak, do not assume you are safe. Weak detection today is not a guarantee tomorrow. The routine above is the same either way.

If your only goal is to fool a detector, stop. That is not what this is for. The watermark exists because platforms want to label model output. Your job is to make sure the text you submit is actually yours.

The evergreen fix, and where to go next

Watermark or no watermark, the advice is the same: treat ChatGPT output as a first draft, make it sound like you, and add what only you know. That is the whole game.

For the full method, start with our pillar guide: Humanize AI Text: What Actually Works. It covers sentence-level editing, verification, and how to keep your meaning intact while changing the surface.

One more thing worth knowing about our own tooling. On 2026-09-28 we scored 709 passages from DeAIze guides not marked generated with our detector at a 30% flag threshold. The mean AI score was 26%, the median 25%, and 239 of 709 passages landed at or above the threshold. That is our own measurement, on our own instrument, on a metadata-selected sample. It is not a false-positive rate, and it is not a claim about any other detector. It does show something useful: even text we did not mark as generated produces a spread of scores, and a flag is a signal to inspect, not a verdict.

If you want to see which sentences in your own draft carry the signal, run it through the detector and read the line-level output before you rewrite a word.

---

Measurement note: figures in this article come from our own detector run on 2026-09-28 — n = 709 reference passages from guides not marked generated versus n = 12 unedited model passages, median AI score 25% and 34% respectively. Reference authorship was not independently verified; passages are not independent documents. Method: /methodology/.