What Is OpenAI textGrain? ChatGPT's EU Text Watermark
The short answer
textGrain is OpenAI’s invisible watermark for AI-generated text. Announced on October 5, 2026, it slightly biases the model’s word choices in a statistical pattern that a detector can recognise, without changing what the text says. OpenAI is adding it to eligible ChatGPT and Codex output in the European Union over the following weeks to comply with the EU AI Act, and API customers anywhere can opt in for select models (off by default). The detector is not public: only approved researchers and expert organisations can apply.
How it works
A language model picks each next word from a ranked list of candidates. textGrain nudges those choices toward a pattern derived from a secret key; over a few hundred words, the pattern becomes statistically unlikely to occur by chance. A detector with the key measures how strongly the pattern is present and reports whether it finds an OpenAI watermark. OpenAI has published a technical report, says it will add detail in the coming weeks, and plans to release the technology as open source.
OpenAI reports no meaningful quality loss. On GPT-6 Astra at max effort, the Artificial Analysis Intelligence Index was 49.57 unwatermarked and 49.76 watermarked; GPQA Diamond 94.44% versus 93.94%; DeepSWE v1.1 72.80% versus 71.68%.
What it detects — and what it doesn’t
| Condition (1% false-positive target) | Detection rate |
|---|---|
| 400-token passage, psychology content | ~95% |
| 200-token passage, psychology content | ~80% |
| Mathematics content | ”Substantially lower” |
| 400 tokens, 10% of words replaced by synonyms | ~66% (from ~92%) |
| 400 tokens, 25% of words replaced | ~17% |
OpenAI is explicit about what a positive result does not mean: it does not measure human contribution, establish ownership or responsibility, identify the user or their prompts, or verify accuracy. A negative result does not prove human authorship — the text may be short, edited, translated or from an unsupported model.
How it compares with other provenance signals
- Google SynthID for text is the closest equivalent; OpenAI says textGrain matched or exceeded it and other approaches in its evaluations.
- Content Credentials (C2PA) record a file’s origin in metadata. OpenAI adds them to supported images; metadata is easily stripped, which is why vendors pair it with watermarks.
- AI text detectors (Pangram, GPTZero, Originality.ai) guess from writing style without any key. They work on any model’s output but produce more false positives.
What it means for you
- ChatGPT users in the EU: your output will carry a watermark you cannot see. It does not tag your identity or conversation.
- Developers: the API opt-in lets you watermark your own product’s output for EU AI Act Article 50 compliance without building it yourself.
- Teachers and editors: do not expect a public “ChatGPT checker” from OpenAI soon. Use version history and detectors — see how to find out if a document is AI generated.
Related: Claude watermark vs SynthID vs C2PA.
Last verified: October 10, 2026.