OpenAI sets out phased text watermarking rollout

OpenAI says the EU AI Act requires generative AI providers to make generated text identifiable in a machine-readable way.[2] Its textGrain technology adds an invisible statistical signal to a model’s word choices, and a detector checks whether a passage contains an OpenAI watermark.[2] API customers worldwide can opt in for select models, with the feature off by default.[2] Separately, eligible ChatGPT and Codex text output will receive invisible watermarks in the EU over the coming weeks, covering eligible users across all plans there.[2]

The ChatGPT and Codex rollout is limited to the EU at launch; OpenAI says text watermarking will not be a global default.[2] The company is also working with cloud partners to make watermarking available for OpenAI model outputs through their services in the coming weeks.[2] Approved researchers and expert organizations can apply starting today, but access will initially be granted case by case to support evaluation and improvement of text provenance.[2] OpenAI says the detector reports whether it finds a watermark without identifying users or revealing prompts or conversations.[2]

At a 1% target false-positive rate, OpenAI says its detector found watermarks in about 80% of 200-token passages and 95% of 400-token passages for content such as psychology.[2] Detection rates were substantially lower for mathematics, where word choice is less flexible.[2] Editing can also weaken detection: replacing 10% of words with synonyms in 400-token passages reduced detection from about 92% to 66%, while replacing 25% reduced it to 17%.[2] OpenAI cautions that ideal-condition performance does not guarantee reliable everyday detection.[2]

OpenAI says a watermark does not measure the amount of human judgment, editing, or creativity in a passage, establish ownership or responsibility, or verify whether text is true, misleading, harmful, or presented in the right context.[2] The absence of a detected watermark also does not prove that a person wrote the text.[2] OpenAI says its image and audio verification tools, including openai.com/verify and the Content Provenance API, will remain publicly accessible.[2] It plans to make textGrain open source so others can build on it.[2]