OpenAI to add textGrain watermark to EU ChatGPT and Codex output

OpenAI says textGrain will mark eligible ChatGPT and Codex text in the European Union over the coming weeks. The measure follows transparency rules under the EU AI Act that took effect on August 2. The company says API customers worldwide can opt in for selected models, while the detector remains restricted initially.

The rules require companies to mark AI-generated content so other systems can identify it. OpenAI describes textGrain as an invisible watermark embedded through changes to model word choices. It is not a visible symbol, and the pattern is intended for detection by a tool using the text and a secret key. OpenAI says the mark travels with text when copied and pasted.

The rollout follows transparency rules

The company says the EU rollout will cover eligible ChatGPT and Codex users across all plans. It has not set out in the supplied material a more precise timetable than the coming weeks. OpenAI says the initial regional approach allows it to learn from use and feedback before making further decisions. It also says text watermarking will not be a global default at launch.

For API customers, OpenAI says opt-in availability begins immediately for selected models, anywhere in the world. The feature is off by default, leaving customers to decide whether to enable it for their own use. OpenAI says it is also working with cloud partners to make watermarking available for model output through their services in the coming weeks. Those partner arrangements and the precise models are not specified in the supplied material.

Detection has limits in practice

OpenAI’s technical report describes a method in which a secret key sorts next-word predictions, shaping output through a series of small choices. Across a passage, those choices can create a pattern a detector can identify from the text and key. The report was co-written with researchers from Pennsylvania and Yale. The method does not place a visible label on the text.

OpenAI reports that replacing 10% of words with synonyms reduced detection from about 92% to 66% in one test. The company also says short passages, mathematical answers and translated text are harder to identify reliably. These findings limit what a positive or negative result can establish. OpenAI says a missing watermark does not establish human authorship, since text may be short, edited, or generated by another system.

Access remains limited for evaluation

OpenAI says approved researchers and expert organisations can apply for detector access, with initial access granted case by case. The tool is intended to report whether it detects an OpenAI watermark. According to the company, it will not identify a user or disclose prompts or conversations. OpenAI says the limits and risks of missed marks and false positives explain why it is not making the detector public at launch.

The company says watermarks can indicate that an OpenAI system generated or processed part of a passage. It cautions that this does not show how much human judgement, editing or creativity contributed. OpenAI also reports no meaningful change in model performance when watermarking is enabled. It has cited benchmark results showing similar performance for marked and unmarked text, while warning that detection is not guaranteed.

OpenAI has not specified which selected models will support API opt-in, or when cloud partner access will begin. The effectiveness of textGrain after editing and across different kinds of text remains a stated limitation. The next steps described by the company are the EU rollout, API opt-in access and case-by-case detector review. Any wider availability will depend on decisions not detailed in the supplied material.