OpenAI says its text watermark detector is free, and not for checking student work

Published 10 October 2026

OpenAI's help centre article on provenance signals gained a new question late on 9 October 2026: who qualifies for text detector access, and are checks billed. The answer is that text detector checks are currently free, that the qualifying uses are academic and research organisations studying text provenance, detection reliability or how people understand provenance results, and that academic integrity applications, such as checking student work for AI use in essays or theses, are currently outside the programme's scope[15].

What changed

Until this edit the article said only that organisations with a qualifying use case could apply for text detector access, reviewed case by case, and that switching watermarking on as an API customer does not grant it. Both statements remain. The new answer adds three things: the checks cost nothing, general business applications such as commercial fraud prevention, authentication and verification, or reselling as part of packaged provenance services, are out of scope, and so is academic integrity. The page carries a modification time of 22:51 UTC on 9 October[15]. The same edit widened the neighbouring answer on the image and audio Content Provenance API, which now names newsroom verification, fact checking, provenance research and regulatory oversight as qualifying work[15].

What it means

By OpenAI's own statement, checking whether a student used ChatGPT is not a use its text detector programme currently admits. The watermark itself is unaffected: the help centre still lists text as textGrain watermarks, EU only, with the API opt-in worldwide, and the tracker row stays Partial with a Gated detector[15]. What to do if a detector flags your text.

What is still not known

OpenAI does not say how many organisations have access, how long review takes, or whether the scope will widen; it says its approach to access may evolve. Which models can be watermarked is still visible only inside the API settings, and OpenAI has not said the EU rollout to ChatGPT and Codex is complete[15]. The tracker.

Sources

Provenance signals in OpenAI-generated content primaryOpenAI Help Center, 2026-10-06, edited 2026-10-07 and 2026-10-09 (the page shows only a relative 'Updated:' line; the absolute times come from an updatedAt timestamp embedded in the page's JSON, read as 22:36 UTC on 7 October and 22:51 UTC on 9 October 2026)

Last verified against primary sources: 2026-10-10