OpenAI plans to add invisible watermarks to eligible ChatGPT and Codex text in the European Union over the coming weeks, while opening optional watermarking to API customers worldwide.
In its Oct. 5 announcement, the company described an EU-only rollout across all plans for its own products. API customers can opt in for select models starting that day, with watermarking remaining off by default. OpenAI framed the phased approach as its response to EU text-provenance rules.
A signal in word choices
The technology, textGrain, creates a statistical signal in the model’s word choices. Its technical report describes adjusting the sampling of tokens, words or parts of words, using randomness derived from a secret key. An entropy budget limits the average sampling randomness removed to create the signal. A detector uses the text and key to look for that pattern.
OpenAI reports no meaningful performance difference with and without watermarking across its Astra benchmarks. It plans to release the technology as open source.
Detection has limits
OpenAI is opening detector-access applications, initially for approved researchers and expert organizations rather than the public. Its evaluations show why interpretation needs care: in a test of 400-token watermarked English responses to questions from the ELI5 dataset, replacing 10% of words with synonyms reduced detection from about 92% to 66%; replacing 25% reduced it to 17%.
The company also says short or more constrained text is harder to detect. A watermark does not establish ownership, accuracy or the amount of human contribution. Conversely, failing to detect one does not prove a passage was written by a person.
