OpenAI brings text watermarks to EU users, with explicit limits on detection

Selected API models can use it worldwide by opt-in; it is off by default there. Access to the text detector initially requires approval. [1]

The mark lives in statistical word choices, rather than hidden characters. During generation, a secret key influences which plausible tokens are selected. A matching detector looks for that pattern across a passage. Short or tightly constrained text offers fewer choices from which to recover a signal. [2][3]

OpenAI’s editing test shows why a negative result needs care: detection fell from about 92% to 66% and then 17% as increasing shares of words were replaced. These are controlled company tests, not a universal measure of whether a person wrote a document. [1]

For editors and teachers, our recommendation is to retain drafts, sources and an account of the writing process. A detector result can prompt inquiry; it cannot fairly substitute for that evidence. OpenAI says a watermark identifies neither the user nor how much human judgment went into the text. [1]