Yes, but only in some places for now. Since October 5, 2026, OpenAI lets API customers turn on text watermarking for select models. It’s off by default. In the coming weeks, OpenAI will add an invisible watermark to text from ChatGPT and Codex for eligible users in the European Union, on every plan. Outside the EU, ChatGPT doesn’t watermark its text yet.
Three things change the answer for you: where you use it, which product you use, and how much the text gets edited afterward.
What did OpenAI change?
OpenAI set out its approach to the EU AI Act’s rules for marking AI-generated text (OpenAI, reported by Unite.AI). The EU rules apply from August 2, 2026.
- API, worldwide: opt-in for select models, off by default. The developer decides.
- ChatGPT and Codex, EU only: watermarking for eligible users on all plans, rolling out over the coming weeks.
- Detector: not public. At first, only approved researchers and expert organizations can apply for access.
OpenAI hasn’t said exactly which users count as eligible, or whether ChatGPT users in the EU can turn it off.
What is a text watermark, and what isn’t it?
It’s a pattern in word choice, not a mark you can see.
When a model writes, it picks each next word from a set of likely options. A watermark nudges those picks in a way only someone with the key can check. Anthropic describes its version as changing the randomness behind word selection, without changing quality or adding tokens (Anthropic).
So the watermark is not a hidden character. If you find odd spaces or zero-width characters in pasted text, those aren’t it.
OpenAI is also clear about what its watermark can’t do. It can’t measure how much a person contributed, prove who owns a text, identify a user or check whether the text is accurate. It can only say whether a passage likely carries OpenAI’s watermark.
How reliable is it?
It works on long, mostly unedited passages and gets weak quickly after that. These are OpenAI’s own numbers, at a 1% false positive rate:
- About 80% of 200-token passages and about 95% of 400-token passages were detected.
- Math was detected much less often.
- Replacing 10% of the words with synonyms dropped detection from about 92% to 66%. Replacing 25% dropped it to 17%.
A 200-token passage is roughly 150 English words. A short reply, a single paragraph or a text you rewrote in your own words gives the detector little to work with.
Does Claude or Gemini watermark text?
Yes, both do, and they started earlier than ChatGPT.
| ChatGPT | Claude | Gemini | |
|---|---|---|---|
| Watermarks text | EU only, rolling out from October 2026; API opt-in | Yes, worldwide, since August 2026 | Yes, since 2024 |
| Method | OpenAI’s own | A version of SynthID-Text | SynthID-Text |
| Who can check it | Approved researchers and organizations | Eligible organizations, such as regulators, media, researchers and schools | Not public |
| Can you turn it off | API: off by default. ChatGPT: not stated | Not stated | Not stated |
Claude. Anthropic adds the watermark to text from newer Claude models and applies it globally. Its detection API is in private preview for organizations eligible under EU law. Anthropic says the mark is weak on short text and on factual passages with few word choices. It can’t tell text Claude wrote from text Claude edited, and it can’t identify a human author. On editing, it says: “Light editing probably won’t remove the watermark completely; a complete rewrite where every word is replaced will” (Anthropic).
Gemini. Google DeepMind has used SynthID to watermark text from the Gemini app and web experience since 2024 (Google DeepMind).
What does this mean if you write with AI?
Less than the headlines suggest.
Text you write yourself carries no watermark. Asking AI to fix grammar in your draft barely changes it. What carries the watermark is text a model generated and you pasted as is, in long stretches.
The people who could check the watermark are researchers, regulators and a few organizations. Your teacher, client or editor can’t run it today. If someone questions your writing now, it’ll most likely be because of an AI detector score, which is a different thing. I wrote about why detectors flag human writing in Why Does My Writing Get Flagged as AI When I Wrote It Myself?
My own view: the watermark isn’t the important part. It answers “did a model produce these words?” That isn’t the question a reader cares about. Readers care whether the piece is worth their time, whether it’s true, whether it says something they didn’t know. Arguing over whether a text was written by hand or with AI misses that. Good writing with AI looks the same as good writing without it. It starts from what you know and think, and you keep working on it until every sentence says what you mean.
Where StashBase fits
If you write with AI, the useful habit is the same with or without a watermark: start from your own notes and sources, and revise until the text is yours.
StashBase is a local writing workspace for that. You keep your notes, sources and drafts as ordinary files in one project, and draft and revise with Claude Code, Codex or Antigravity beside them. When the agent revises your prose, the changes show up in the document as suggestions you accept or reject one at a time, so you decide what stays. I wrote more about this way of working in How to Write with AI Without Losing Your Voice.
Frequently asked questions
Can I turn off the ChatGPT watermark?
In the API, watermarking is off unless the developer turns it on. For ChatGPT and Codex in the EU, OpenAI hasn't described a setting to turn it off.
Can teachers detect the ChatGPT watermark?
Not at launch. OpenAI's detector isn't public; it's limited to approved researchers and expert organizations that apply for access. Schools still mostly use AI detectors like Turnitin and GPTZero, which work differently.
Is the watermark a hidden character in the text?
No. It's a statistical pattern in which words the model picks. You can't see it, and it isn't a special space or zero-width character.
Does Claude watermark text?
Yes. Since August 2026, Anthropic adds an invisible watermark to text from newer Claude models worldwide, using a version of Google DeepMind's SynthID-Text.