OpenAI is rolling out an invisible watermarking system for text generated by its ChatGPT and Codex models within the European Union. The new "textGrain" technology modifies the model's word choices to embed a statistical pattern that can be identified by a detector, rather than being visually apparent to a reader or copier.
The company confirmed that the invisible watermark will be applied to eligible ChatGPT and Codex text outputs in the EU "over the coming weeks." While this feature is not yet a global default, API developers worldwide can opt-in to watermarking for supported models, although it remains disabled by default. OpenAI is also accepting applications for access to its watermark detector, with initial access limited to approved researchers and expert organizations.
OpenAI acknowledges that the text watermarking system is not foolproof, noting that even moderate editing can significantly reduce its effectiveness. Internal evaluations showed that replacing 10% of words with synonyms in 400-token passages decreased detection rates from approximately 92% to 66%. This rate dropped further to 17% when 25% of words were replaced.
The company's tests, conducted at a 1% false-positive target, indicated that the watermark was detected in about 80% of 200-token psychology responses, improving to roughly 95% for 400-token texts. Detection proved less reliable for subjects like mathematics, where the AI model has less lexical flexibility.
OpenAI explicitly warns that the absence of a detected watermark does not confirm human authorship. Factors such as text being too short, edited, or translated can hinder reliable detection. Furthermore, a detected watermark does not reveal the generator's identity, account, prompt, or conversation, nor can it quantify the extent of human involvement in the final work.
Despite these limitations, OpenAI states that enabling textGrain does not meaningfully impact the quality of its GPT-6 Astra model, with benchmark results remaining largely consistent.
