Where textGrain intervenes in the act of writing
A language model generates text by repeatedly choosing the next word from a probability distribution. A statistical watermark does not crudely alter that distribution; it replaces part of the randomness with pseudo-random values computed from a secret key and the preceding words. The vocabulary is partitioned into blocks, and the system leans slightly towards certain blocks while the probabilities of words inside the chosen block remain as they were.
The consequence is that each individual choice looks entirely ordinary, yet when hundreds of choices sit side by side they form a pattern only a detector holding the secret key can read. That is why the technique works on long text and weakens very quickly on short text.
"A secret key sorts next-word predictions, with hundreds of these nudges together enabling detector identification." — OpenAI technical documentation
Who it applies to, who gets to choose, and why now
According to technology news outlets, the watermark will be switched on automatically over the coming weeks for ChatGPT and Codex for users in the European Union, across all subscription tiers. For developers using the programming interface, the feature is off by default worldwide and must be enabled in organisation or project settings.















