ChatGPT Text Watermarks Explained: What textGrain Can—and Cannot—Prove

Does ChatGPT watermark text now?

OpenAI's October 5, 2026 announcement describes a phased rollout for eligible ChatGPT and Codex text in the EU over the following weeks. It is not a worldwide default at launch. API customers globally can opt in for selected models; the API default remains off. Text detector access initially requires approval for researchers and expert organizations.

These are three separate questions: whether a product creates a watermark, whether a particular passage contains one, and whether you have access to a detector. An announcement about the first does not answer the other two. We have not verified rollout for individual accounts.

Reviewed October 7, 2026, against the announcement, Help Center, and technical report. This article explains published evidence; it does not report an independent test of the restricted text detector.

How textGrain works, in plain English

The textGrain technical report describes a method that links token selection to randomness derived from a secret key and the preceding context. An entropy budget controls how much sampling variation the method can trade for that signal. The detector reconstructs the relevant pattern using matching settings and the key, then tests how the observed tokens align with it.

Think of this as evidence accumulated across choices. A particular word does not carry a visible stamp saying “AI.” Looking at one sentence and deciding that it sounds polished is a different activity from testing the keyed pattern.

The OpenAI provenance FAQ makes the practical distinction explicit: textGrain does not insert hidden characters, invisible spaces, or special punctuation. Ordinary AI classifiers analyze text after generation; watermark detection looks for an embedded signal.

What can—and cannot—a watermark result prove?

OpenAI's provenance FAQ says a watermark can support an inference that its models generated or processed content. It does not establish who wrote it, how much a person contributed, who owns it, or whether it is accurate. Our editorial interpretation is to keep the report as narrow as the measurement:

Imagine an editor receives a draft containing original reporting, a quoted passage, and an AI-assisted rewrite of a conclusion. Before interpreting any document-level label, the editor needs to know what text was submitted and what the tool actually measured. A label cannot reconstruct that editing history.

Why detection rates need context

In its announcement, OpenAI reports approximately 80% detection for 200-token passages and 95% for 400-token passages in content such as psychology, at a target 1% false-positive rate. Mathematics performed substantially worse. These are evaluation results, not guarantees for your document.

The technical report also distinguishes theoretical calibration from deployment: its error-rate result relies on stated assumptions, and a fixed key and numerical implementation require empirical checks.

A 1% false-positive rate does not mean that every positive result has a 99% chance of being correct. For an illustrative calculation, a test with that rate applied to 10,000 unwatermarked passages would flag about 100 on average. To interpret the positive results in a mixed collection, you also need the prevalence of watermarked text and the test's detection rate. This example explains the arithmetic; it is not a measurement of textGrain on a real collection.

OpenAI also reports that edits can weaken detection. Treat that as a limit on interpreting results, not a dependable recipe for making a text undetectable. A before-and-after score from an unrelated classifier does not validate watermark removal.

What writers and editors should do

Our recommendation is to keep a small evidence trail alongside the draft. It is more useful during an editorial review than a screenshot of a score with no explanation of the test.

1. 01
   Keep the original and the revision.

   Save the source material, your draft, and substantive edits. Note where an assistant helped with structure, wording, or research so that those roles can be explained later.

2. 02
   Name the actual check.

   Record the tool, date, submitted passage, and type of result. Keep “classifier score,” “character scan,” and “watermark detection” as separate labels in your notes.

3. 03
   Review claims before style.

   Check facts, quotations, attribution, and qualifications against the sources. Then use a text comparison to see whether a rewrite changed the substance.

4. 04
   Resolve questions with context.

   If a review raises concerns, examine the drafting history and relevant publication rules. Ask what additional evidence would answer the specific question before drawing a conclusion.

UltraHumanizer can help refine wording and review changes. We do not present its humanizer, AI detector, or text cleaner as an authorized textGrain detector or as a guarantee of watermark removal. Choose an editing tool for the writing task you can actually inspect.

For another example of reading AI claims carefully, see our explanation of OpenAI's 722 math manuscripts and their verification status.