Build1 publisher2 min readPublished
OpenAI's textGrain text detector goes to approved researchers first
OpenAI is giving its textGrain watermark detector for ChatGPT and Codex text only to approved researchers and expert organizations, case by case. Teams reviewing outside documents have no public checker and no date for one, so their AI-content policies have to rest on their own records.
The Engineer · Build desk
Drafted by a language model from the sources cited here and checked against its claim ledger before publication. How we use AISend a correction
What happened
- OpenAI's API customers can opt in to receive text outputs that carry the textGrain watermark.
- The detector reports only whether a passage contains an OpenAI watermark, and it reveals neither the prompt behind the text nor the person who used the product.
- Short passages are hard to detect, and rewriting, editing, or translating a passage can lower the reliability of a detection result.
- OpenAI presents the phased rollout as part of a layered provenance strategy and as an effort to meet EU AI Act requirements.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- exposure Policies that read a negative result as human authorship will misclassify text from other providers, human drafts written with AI help, and anything generated without watermarking enabled.
- decision API customers who want their own copy verifiable later have to opt in at generation time, and each round of heavy editing before publication reduces what that opt-in is worth.
- cost Proving how a document was made falls back on records a team keeps itself: whether AI was permitted, which tools were used, what human review happened, and whether claims were checked.
An outside reviewer gets a usable textGrain result only when three conditions hold together [19]. First, the text was generated with watermarking enabled [10]. Second, it reached the reviewer without being condensed, paraphrased or translated, since any of those can change the detection result [9]. Third, the reviewer can actually run the detector. OpenAI is granting that access case by case to approved researchers and expert organizations [4]. A compliance team outside that group fails the third condition before the first two matter [19].
According to a dev.to summary of OpenAI's EU text provenance announcement, OpenAI calls textGrain a durable watermark for eligible text outputs [11]. The same write-up says text watermarks are less durable than provenance signals for images and audio [13]. It also says language variation alone can affect detection rates [9]. Given that, I think gating the detector is the right call. OpenAI says the restricted phase is for evaluation and improvement [4]. It also wants feedback on how results should be communicated [14]. The failure it is guarding against is a user treating an uncertain result as definitive proof, in either direction [17].
The design also assumes the watermark will sometimes be the only provenance signal left [8]. OpenAI's stack combines four layers. C2PA metadata works only while it stays attached to the content. Durable watermarks are meant to persist when that metadata is removed. Verification tools and expert evaluation make up the rest [8]. The write-up lists verification tools twice [8], so that layer at least has redundancy.
The post does not include a date for wider access [15]. In my view, a review process should be built on the assumption that the detector stays out of reach, with any later access treated as an addition. The dev.to author goes further. Even with access, the author argues, a detector result should not be the sole basis for a compliance decision, a contractual dispute, or an editorial judgment [12].
What to watch
- Any OpenAI date or eligibility criteria for opening textGrain detection beyond approved researchers and expert organizations.
- Published detection rates from the approved testers for short, edited, and translated passages.
- Whether OpenAI specifies which ChatGPT and Codex outputs count as eligible for watermarking outside the API opt-in.