On October 5, 2026, OpenAI announced a text provenance signal called textGrain. Select API customers can opt in globally; eligible ChatGPT and Codex text output in the EU is planned for the coming weeks. This does not mean every AI-generated paragraph already carries a detectable mark.
- How the rollout actually works
- Detection is probabilistic, not proof
- What a result cannot establish
- A more useful provenance workflow
How the rollout actually works
The API option is off by default and limited to selected models. OpenAI says detector access begins with approved researchers and expert organizations, not a public checker. The EU rollout is prospective as of October 7. Image and audio verification are separate systems.
Detection is probabilistic, not proof
The signal is statistical and invisible in ordinary reading. OpenAI reports roughly 80% detection for 200-token and 95% for 400-token passages in one psychology-content test at a 1% target false-positive rate. These figures are conditional, not universal accuracy promises; edited text and other subjects can perform worse.
What a result cannot establish
A positive result does not prove who wrote a passage, whether it is true, or who owns it. A negative result does not prove a person wrote it: the text might come from an unmarked model, be too short, or have been substantially edited. A single detector result should not decide a consequential accusation.
A more useful provenance workflow
Publishers, schools and teams should keep drafts, source links, revision history and clear disclosure when AI contributes. Use watermark detection only alongside those records and human review. This gives readers more context than a bare AI-or-human label.
A new invisible signal for some AI-written text is rolling out. Its detection limits matter more than the headline.
VerdictLab editorial note
Original sources
This VerdictLab report is an original summary and analysis based on the reporting below.
How this was created: VerdictLab selected and checked the cited reporting, prepared an original summary with AI assistance, and translated it for accessibility. We link to the original publishers and do not reproduce their articles.