OpenAI Adds Invisible Watermarks to ChatGPT Text in the EU
OpenAI said on October 5, 2026 that it will add an invisible watermark to eligible ChatGPT and Codex text output in the European Union. The technology, called textGrain, does not insert anything into the text. It nudges the model's word choices so the passage carries a statistical pattern that a detector can pick up later.
The change is being rolled out over the coming weeks, applies only in the EU, and comes with limits OpenAI itself spells out. Here is what is confirmed, what the detection numbers show, and what a watermark can and cannot tell you.
⚡ Quick facts
- Company: OpenAI
- Announced: October 5, 2026
- Technology: textGrain, a statistical text watermark
- Where: ChatGPT and Codex in the EU, rolling out over the coming weeks, all plans
- API: opt-in worldwide for supported models, off by default
- Detector: applications open, limited to approved researchers and expert organizations
- Cost: no pricing change reported
What OpenAI Announced
On October 5, 2026, OpenAI announced that it will watermark text from ChatGPT and Codex for users in the EU. In the wording reported by BleepingComputer, the company said it would add an invisible watermark to eligible text output in the European Union over the coming weeks. The same day, developers using the OpenAI API anywhere in the world gained the option to turn the watermark on for supported models.
ITdaily and other outlets tie the move to the transparency obligations in the EU AI Act, which expect providers to mark AI-generated content in a machine-readable way. The reports I reviewed do not quote OpenAI citing the law directly, so treat that link as the outlets' reading rather than OpenAI's stated reason.
How textGrain Works
A language model picks each next word from a ranked set of likely options. textGrain steers those choices in a consistent, hard-to-notice way, so the finished passage contains a pattern that would be very unlikely to appear by chance. A detector that knows the pattern can score a passage and report whether it matches.
Because the signal lives in the wording, there are no invisible characters, extra spaces or unusual punctuation to remove. Copying a response into another document does not wipe it out. It is also why heavy rewriting can damage it.
What the Reports Confirm
- Scope: eligible ChatGPT and Codex text output in the EU, over the coming weeks, on all plans. It is not a global default.
- API: opt-in for supported models worldwide, disabled by default.
- Detector: OpenAI is accepting applications, with access limited at first to approved researchers and expert organizations.
- Quality: OpenAI says GPT-6 Astra benchmark results stay broadly similar with textGrain enabled.
The details above come from reporting on OpenAI's announcement, mainly BleepingComputer, TechRepublic and ITdaily. OpenAI's own post was not available to read directly when this article was written, so exact wording may differ slightly.
How Reliable Is Detection?
OpenAI published test figures that show both the promise and the weakness of the approach:
| Test | Result reported |
|---|---|
| Unedited 400-token passage | About 92% detected |
| 10% of words swapped for synonyms | About 66% detected |
| 25% of words swapped for synonyms | About 17% detected |
| 200-token psychology responses (1% false-positive target) | About 80% detected |
| 400-token psychology responses (same target) | About 95% detected |
Longer text is easier to detect. Subjects like mathematics are harder, because the model has less freedom to vary its wording. OpenAI also warned that “the absence of a detected watermark does not prove human authorship.” Short, edited or translated text may be too altered for a reliable result.
Why It Matters
Until now, tools that guess whether a text was written by AI have been unreliable and have wrongly flagged people. A watermark applied at generation time is a firmer signal, but only for text produced by a system that applies it. Anyone who paraphrases the output, uses another model, or runs ChatGPT outside the EU is not covered.
The watermark also does not identify a person. According to the reports, a detected watermark does not reveal the user, account, prompt or conversation, or how much of a piece a human wrote or edited. It says only that the wording matches a pattern.
Who Is Affected
EU users of ChatGPT and Codex are the first group, though they will not see any change on screen. Researchers and organizations that want to check text can apply for detector access. Developers using the API can decide whether to enable the watermark in their own products. Readers elsewhere, including in India and the US, are not affected unless an app they use has switched it on.
Cost and Availability
No price change was reported. The watermark applies on all ChatGPT plans in the EU, and API use is optional for supported models. OpenAI has not published a date by which every EU output will be covered, only that the rollout will take place over the coming weeks.
What Happens Next
The EU rollout is expected over the coming weeks from October 5, 2026, and detector access will widen only as OpenAI approves applicants. Reports also say OpenAI plans to release the technology as open source so other parties can build on it; check OpenAI's announcement for the timing before relying on that.
Frequently asked questions
What is OpenAI's textGrain watermark?
textGrain is the invisible text watermark OpenAI announced on October 5, 2026. It slightly shifts the model's word choices to leave a statistical pattern that a detector can later find. Nothing is added to the text itself, so there are no hidden characters to strip out.
Who gets watermarked text from ChatGPT?
OpenAI said it will add the watermark to eligible ChatGPT and Codex text output in the European Union over the coming weeks, on all plans. It is not a global default. Outside the EU, only API developers who opt in will produce watermarked text.
Can I tell if a text came from ChatGPT using this?
Not yet for most people. OpenAI is taking applications for its detector, but access starts with approved researchers and expert organizations. OpenAI also warns that failing to detect a watermark does not prove a person wrote the text.
Does editing or paraphrasing remove the watermark?
Heavy edits weaken it. In OpenAI's tests on 400-token passages, swapping 10% of words for synonyms cut detection from about 92% to 66%, and swapping 25% cut it to 17%.
Does the watermark change the quality of ChatGPT answers?
OpenAI says benchmark results for GPT-6 Astra stay broadly similar with textGrain switched on. That is OpenAI's own claim and has not been independently tested in the reports reviewed.
Will this affect users in India or the US?
Not by default. The rollout OpenAI described covers the EU only. Developers anywhere can switch the watermark on through the API for supported models, but it is off unless they choose it.
Comments