OpenAI has detailed a phased text-watermarking rollout designed to meet transparency requirements under the EU AI Act. Starting October 5, API customers worldwide can opt in to watermarking for select models, while OpenAI plans to add invisible watermarks to eligible ChatGPT and Codex text output in the European Union over the coming weeks. The company is not making text watermarking a global default at launch.
The system, called textGrain, changes the statistical pattern of model word choices rather than inserting visible labels or hidden characters. A detector then looks for that signal in a passage. OpenAI says it plans to open-source the technology, but detector access will initially be limited to approved researchers and expert organizations because the signal is imperfect and can be weakened by editing.
OpenAI’s own evaluations illustrate the limitation. At a target false-positive rate of 1%, the company says detection reached about 80% for 200-token passages in some content and about 95% for 400-token passages, while constrained domains such as mathematics performed worse. In another test, replacing 10% of words with synonyms reduced detection from roughly 92% to 66%, and replacing 25% reduced it to 17%. Those are vendor evaluations, not independent validation.
The company also stresses that a detected watermark does not establish ownership, identify the user, measure the amount of human contribution or verify whether the text is accurate. Conversely, failure to detect a watermark does not prove human authorship. OpenAI says these limits are why it is restricting initial detector access and treating watermarking as one part of a broader provenance strategy.
Independent reporting from The Verge confirms the regional rollout, API opt-in and restricted detector access. The policy significance is substantial: a major model provider is adapting text generation behavior specifically to an external legal transparency requirement, while openly documenting the technical fragility of the detection mechanism. For developers and organizations operating in Europe, provenance is therefore becoming a deployment and compliance consideration rather than merely a research topic. The unresolved question is how reliable such signals remain after ordinary editing, translation and mixed human-AI authorship.