What is textGrain?
textGrain is OpenAI’s method for embedding an invisible watermark in AI-generated text. During generation, it statistically adjusts the selection of words or word pieces (tokens). Across a longer passage, these choices form a signal that a suitable detector can examine using a secret key.
The watermark does not consist of hidden characters, extra spaces or unusual punctuation. A character scanner therefore does not test this statistical signal. textGrain provides a provenance signal for supported OpenAI outputs; it is not a general detector for every kind of AI-generated text.
Practical example
An editorial team receives a draft created with help from an AI tool. It checks the facts, sources and editorial approval. An available provenance signal may complement that review. It does not replace it or reliably identify either the human contribution or the person behind the text.
What to watch out for
- False positives and missed watermarks are possible. Short or heavily edited passages are harder to detect.
- Failure to detect a watermark does not prove that a human wrote the text.
- A watermark does not confirm factual accuracy or ownership of the text.
- Availability and detector access may change. OpenAI currently describes access to its text detector as restricted.
Context and availability
OpenAI introduced textGrain on 5 October 2026 in the context of EU transparency rules. Its announced rollout covers eligible ChatGPT and Codex text outputs in the EU; API customers can optionally enable watermarking for supported models. Coverage depends on the product, model and creation date. This EU-related rollout is not a universal rule for every country or AI provider. See also: EU AI Act.
Sources checked: 27 October 2026.