NewsOpenAI2 min read
ChatGPT Will Watermark Its Text in the EU
To comply with the EU AI Act, OpenAI will add an invisible watermark to ChatGPT and Codex text in the EU. How it works and what it does not prove.
Key points
- In the coming weeks, ChatGPT and Codex text will carry an invisible watermark, in the EU only and on all plans.
- It is a statistical signal in word choice: you cannot see it, but a detector can find it.
- The detector is not public: only approved researchers and organisations can apply for access.
- Editing the text weakens it a lot, and not finding it does not prove a person wrote the text.
On 5 October OpenAI explained how it will meet one of the obligations of the EU AI Act: making AI-generated text identifiable by machines. Its answer is an invisible watermark in what ChatGPT and Codex write.
Who it affects
- ChatGPT and Codex in the EU: over the coming weeks, on all plans. It will not be a global default outside the EU.
- API: opt-in worldwide for selected models since 5 October, off by default.
- Images and audio: nothing changes; they already had their own provenance signals.
How it works
The technology is called textGrain. While generating text, the model tilts its word choice very slightly following a statistical pattern. A person notices nothing, but a detector that knows the pattern can find it. According to OpenAI it barely affects answer quality, and it plans to release the technique as open source.
Its own numbers show the limits:
- At a 1% false-positive rate, it detects about 80% of 200-token passages (around 150 words) and 95% of 400-token ones, though far fewer in maths.
- Editing largely removes it: swapping one word in ten for synonyms drops detection from 92% to 66%; one in four, to 17%.
Who can check it
The detector is not public. Since 5 October, researchers and expert organisations can apply for access, approved case by case. The detector does not identify the user or reveal their conversations.
OpenAI stresses what the watermark does not tell you: who the author is, whether the text is true, or how much a person contributed. And not detecting it does not prove a human wrote the text.
What it means for you
- If you use ChatGPT in the EU, your text will carry this mark, but you will not see it and it will not change how it writes.
- If you are a student, this is not an “AI homework detector”: your teacher will not have access, and edited text barely keeps the signal. The honest approach is still to say when and how you used AI. See writing better with AI.
- If you teach, do not trust any tool that promises to detect AI text with certainty: OpenAI itself explains why that is not possible.
