OpenAI вводит маркировку текста — детектор пока для исследователейOpenAI starts text watermarking, with detector access limited to researchers

On October 5, OpenAI announced a phased introduction of textGrain: optional watermarking for selected API models worldwide, followed by eligible ChatGPT and Codex text outputs in the EU over the coming weeks. The API setting remains off by default. Detector applications initially target researchers and expert organizations, rather than a public text-checking service. In one company evaluation, replacing a quarter of the words with synonyms reduced detection from roughly 92% to 17%. Those figures describe a particular experiment, not every document.

The help article describes a statistical pattern in word and token choices. The watermark does not insert invisible characters, spaces or unusual punctuation. Provenance signals also do not establish accuracy, lawful use or the creator's identity. Results need interpretation alongside product coverage and subsequent transformations. Image and audio verification is a separate capability; its availability should not be mistaken for general access to a prose detector.

The technical report explains the tradeoff between signal strength and generation randomness. The method couples token choices to keyed randomness, using an entropy-loss budget. Vocabulary blocks reduce the computation while preserving relative probabilities within each block. Detection requires the text and key, without knowing the generation budget. This describes the mechanism and the assumptions behind its mathematics; it is not independent authentication of an arbitrary document.

The European Commission's transparency code overview distinguishes providers' machine-readable marking responsibilities from deployers' disclosure responsibilities for certain content. Technical solutions must account for format limitations and feasible reliability. A company's implementation announcement therefore cannot settle every compliance question for a particular publisher or organization. The announcement gives readers a product decision to examine, rather than a complete rulebook for their own situation.

Our editorial reading is that origin, accuracy and responsibility deserve separate questions. Imagine a draft that a person reorganizes, checks and shortens. A statistical detector looks for a technical trace; the editor takes responsibility for the claims and publication. Those inquiries overlap but call for different evidence. Keeping the original draft, a record of the tool used and a history of meaningful revisions would make a dispute about assistance easier to examine. One isolated indicator provides much less context.

Why it matters

Text provenance now has another proposed signal, whose interpretation needs the conditions of the check. Readers should still ask what supports a particular assertion. Publishers have an additional practical question: whether their account of the writing process lets a person understand the contribution of tools and the judgment of humans. A watermark can inform that conversation; the surrounding documentation gives it meaning.