What Is TextGrain?
TextGrain is a text watermarking method described in an OpenAI technical report dated October 5, 2026. It creates a statistical relationship between generated tokens and secret-key-derived randomness. It is not a string of invisible characters appended to a response.
From token choices to statistical evidence
- PredictThe language model assigns probabilities to possible next tokens.
- CoupleA secret key and context define vocabulary blocks and pseudorandom costs. Optimal transport couples the choices under an entropy-loss budget.
- SampleA block is selected, then a token is sampled while retaining the original relative probabilities inside that block.
- TestA matched detector reconstructs the keyed quantities from the observed sequence and evaluates statistical evidence.
Our simplified explanation of sections 2 and 3 of the OpenAI technical report, not a production implementation.
What the entropy budget means
The report uses an entropy budget to limit average sampling randomness lost in exchange for a detectable relationship. Its guarantees depend on assumptions and specified marginals. An average-over-randomness claim is not a promise that every fixed key, response or numerical solver run behaves identically.
Detection is keyed
A detector needs the matching secret key and technical configuration. The paper's detection reconstruction does not need the generating model's probabilities or entropy budget. Having source code alone would not reveal the secret key used for another system's production responses.
Four concepts that are often confused
| Concept | Where the signal lives |
|---|---|
| Statistical text watermark | Relationships in the generated token sequence. |
| Invisible Unicode character | A particular code point in the string. |
| File metadata | Information stored around content in a file or container. |
| AI classifier | A model's learned judgment about the text, not a keyed signal. |
Practical limits
Text length, matching configuration, repeated contexts, editing and threshold choice matter. The report's mathematical construction does not certify every deployed detector or every edited text. Continue to detection limits, open-source status or ChatGPT rollout questions.
Sources and verification
Official announcement, access guide and technical report reviewed October 6, 2026. Availability below distinguishes announced rollout from verified access. This site is independent and does not offer a TextGrain detector.