Overview
- OpenAI said on Monday, October 5, 2026 that it will roll out textGrain watermarks to eligible ChatGPT and Codex outputs in the European Union over the coming weeks and let API customers worldwide opt in to watermarking for select models with the API default set to off.
- The company opened applications for a watermark detector but will grant access only to approved researchers and expert organizations rather than releasing the tool publicly at launch.
- TextGrain embeds a hidden statistical signal in a model’s word choices so a detector can spot the pattern, and OpenAI reports detection of about 80% for 200-token passages and about 95% for 400-token passages at a 1% false-positive target.
- OpenAI’s tests show normal edits, synonym swaps, translation, short passages, math and code greatly weaken detection and the company warns a detected watermark only indicates OpenAI generated or processed part of a passage and a missing watermark does not prove human authorship.
- The move implements the EU AI Act’s Article 50 requirement and differs from some rivals by limiting automatic watermarking to the EU; developers and organizations will face new choices about when to enable marking and how to interpret imperfect detector results, which raises operational and fairness concerns for people who heavily edit AI text or use AI to polish second-language writing.