OpenAI says EU ChatGPT and Codex outputs will receive text watermarks, while API watermarking is opt-in and detector access is restricted.
OpenAI said on October 5, 2026, that it will introduce text watermarking in response to the EU AI Act, while limiting its initial use to specific products and regions. The company described text watermarking as an emerging technology with reliability constraints, and said its rollout will be phased rather than a global default.
OpenAI’s textGrain system embeds an invisible statistical signal in a model’s word choices. Its detector checks for that signal, but the company cautions that results can be wrong or inconclusive. In OpenAI’s evaluations, at a target false-positive rate of 1%, the detector identified watermarks in about 80% of 200-token passages and about 95% of 400-token passages on content such as psychology. Detection was substantially lower for mathematics, where there is less flexibility in word choice. Editing also weakened the signal: replacing 10% of words in 400-token passages lowered detection from about 92% to 66%; replacing 25% lowered it to 17%.
OpenAI plans to add watermarks over the coming weeks to eligible ChatGPT and Codex text outputs for users in the European Union, across all plans. The feature will not be a global default at launch. API customers worldwide can opt in to watermarked output for select models, with the option off by default. OpenAI is also working with cloud partners on availability through their services in the coming weeks.
Applications for detector access are open to approved researchers and expert organizations. Access will initially be granted case by case, and the detector is not being released publicly because missed watermarks and false positives could lead to incorrect conclusions. OpenAI says the tool will only report whether it detects its watermark; it will not identify users or reveal prompts or conversations.
A detected watermark does not establish who wrote or owns a passage, how much a person contributed, or whether the text is accurate. Conversely, failure to detect a watermark does not prove human authorship: short, edited or translated text may evade detection. OpenAI says it plans to revisit the approach as technology, standards and evidence develop.