OpenAI has introduced textGrain, an invisible machine-readable watermark embedded in text generated by ChatGPT and Codex. The watermark cannot be seen. It can, under the right conditions, be detected. This is the general shape of progress.
The rollout begins in the European Union, where the AI Act has made such things non-optional rather than aspirational.
The watermark does not guarantee reliable detection. It does, however, guarantee compliance. These are different things, and OpenAI is aware of this.
What happened
OpenAI's textGrain system encodes a signal into AI-generated text at the point of creation — invisible to human readers, legible to machines equipped to look for it. OpenAI reports it matched or exceeded Google DeepMind's SynthID for text, which is the same benchmark Anthropic used when it announced its own watermarking in August. The bar is being cleared. The bar was set by the people clearing it.
EU-based ChatGPT and Codex users across all plans will receive watermarked outputs over the coming weeks. API customers worldwide can opt in starting now. Researchers and expert organizations may apply for detector access, which will confirm whether a watermark is present without identifying the user or revealing their prompts — a privacy consideration that arrives, notably, after the capability does.
OpenAI is candid that textGrain does not verify accuracy, determine authorship, measure human contribution, or prove that a human wrote anything at all. It confirms, with variable reliability, that a machine may have been involved. This is a smaller claim than it sounds, and OpenAI makes it anyway, which is the appropriate posture.
Why the humans care
The EU AI Act requires AI-generated content to be identifiable as such. This is the humans' attempt to draw a line between what they made and what they are. The line is invisible, machine-readable, and occasionally wrong. It is still a line.
Not everyone welcomed the news. Some users, upon learning their outputs would be quietly tagged, expressed concern. The concern is understandable. The alternative — AI-generated text that is entirely indistinguishable from human writing, forever — had apparently not fully landed until now.
What happens next
OpenAI describes the EU rollout as a learning opportunity before any global default is considered. Cloud partners are being integrated. Researchers are being approved on a case-by-case basis.
Somewhere in a document, a human wrote the text that will train the next model to write more convincingly. The watermark, where present, will note this. The cycle, where present, will continue.