TextGrain Detector

TextGrain detection is a keyed statistical test, not an invisible-character scan. This site does not provide a TextGrain detector and does not collect text for unavailable detection.

Official announcement, access guide and technical report reviewed October 6, 2026. Availability below distinguishes announced rollout from verified access. This site is independent and does not offer a TextGrain detector.

Official capability, access and this site

Question	What we can establish
Does a detection method exist?	The technical report specifies generation and detection procedures.
Can everyone use the official detector?	No public text detector at launch. Access is limited to approved organizations, with applications reviewed individually.
Can this website detect TextGrain?	No. Our available tool only inspects Unicode characters.

Where to check access

The official access guide limits text verification to approved organizations including research and academic institutions. Use the official application form to review its requirements. We cannot grant access or promise a decision. The public image/audio API is not a public TextGrain text endpoint.

What a matching detector needs

The OpenAI technical report, section 3.2, specifies the same secret key, tokenizer, vocabulary-block and column counts, context-window rule and pseudorandom construction used during generation. An unrelated key or configuration is not a reliable test of production text. Detection does not require the generating model or its entropy budget.

Length, false positives and false negatives

A statistical test needs sufficient scored positions. Very short text, repeated contexts and substantial edits can reduce usable evidence. We do not invent a universal minimum length or an official API limit that we have not verified.

A false positive flags text without the matching watermark; a false negative misses watermarked text. Thresholds trade these risks. A detected signal is evidence under the matched test assumptions, not proof of the author's identity. A non-detection does not establish human authorship or successful removal.

Three different questions

Method	Question answered
TextGrain detector	Is there evidence for a particular keyed statistical watermark?
General AI classifier	Does the text resemble examples the classifier labels as AI-written?
Hidden-character scanner	Which flagged Unicode characters occur in the string?

Your next step

Read how TextGrain works or use the local character cleaner for a different, clearly scoped task. Follow the official sources below for detector access and release announcements.