ChatGPT en UE: le filigrane textGrain s’efface dès que l’on reformule
OpenAI’s textGrain hides a statistical signal in ChatGPT’s word choices for EU users, yet swapping a quarter of the words leaves it barely detectable.
By
Nathan Lieu
In brief
- OpenAI announced on October 5 that eligible ChatGPT and Codex text in the EU will carry a watermark within weeks.
- The system, textGrain, alters word choices rather than inserting hidden characters.
- Only approved researchers and expert organizations can use the detector at launch.
- Anthropic applies its own text watermark worldwide, unlike OpenAI’s regional default.
For users in the European Union, ChatGPT and Codex text will soon carry an invisible mark that a specialist tool can spot, but only while the wording stays close to what the model produced. OpenAI frames the move as compliance with the EU AI Act, and its own test results show the signal thinning quickly once a person starts rewording. Everywhere else, the feature stays switched off unless a developer turns it on.
A signal hidden in word choice
The technique is called textGrain. Rather than appending anything to the output, it nudges which words the model selects, leaving a statistical pattern a reader cannot see but a detector can measure. According to Search Engine Journal, the company says it is also working with cloud partners so that outputs from its models delivered through their services get the same treatment in the coming weeks.
One practical consequence: OpenAI says textGrain introduces no hidden characters, invisible spaces or unusual punctuation. Tools that strip such artifacts from a document will therefore leave the watermark untouched, because it lives in the vocabulary itself.
The timing follows the regulatory calendar. The transparency rules in Article 50 of the AI Act started applying on August 2, 2026, and the European Commission says systems already on the market before then have until December 2 to satisfy the marking and detection requirement. Outside the bloc, API customers anywhere can opt in for select models.
How editing erodes detection
OpenAI measured its detector on 400-token English passages, then swapped words for synonyms.
Length and subject matter matter too. With the false positive rate set at 1%, the detector spotted the mark in roughly 80% of 200-token passages and about 95% of 400-token passages on topics such as psychology, and far less often on math content.
| Condition | Effect on detection |
|---|---|
| Shorter output | Fewer words to choose from, so the mark is harder to embed |
| Translated text | Lower success than for longer original text, per Ars Technica |
| Rigid vocabulary such as math | Much lower rates than flexible prose |
What a detector result proves
Even a clean hit answers a narrow question. OpenAI sets out several limits:
- A match does not reveal who the user was, nor their prompts or chats.
- No watermark found does not establish that a human wrote the text.
- The detector is not public at launch; approved researchers and expert organizations get access first, and others can apply through an approval process over time.
On quality, OpenAI says benchmarks for Astra, its latest model, show no meaningful difference with the watermark on or off.
Two companies, two maps
Anthropic took a different route in August, when it began marking text to meet the same law. As PCMag reports, its watermark runs globally because the company says it has no durable way to limit it by region. OpenAI argues the opposite: confining the default to the EU leaves room to learn from real-world use and feedback, and it plans to widen detector access once results can be interpreted responsibly.
Featured image. Source: Pexels. Credit: Thirdman. License: Pexels License.
