Sometimes—but not reliably for every passage. OpenAI’s textGrain watermark can help identify eligible OpenAI-generated text when its signal is present and detectable. It is not applied to all ChatGPT output, the detector is initially restricted to approved researchers and expert organizations, and OpenAI’s own tests found misses—especially on shorter or constrained text and after editing. A result is a limited provenance clue, not proof of who wrote a passage or how much a person contributed.
What an AI text watermark detects
OpenAI announced textGrain on October 5, 2026. It embeds an invisible statistical pattern by subtly adjusting the model’s random choices among possible words or word pieces. A detector then checks for that pattern. The watermark is not a visible label or a hidden character in the text.
As an Amazon Associate I earn from qualifying purchases.
This is different from a third-party AI-writing classifier. A classifier looks for statistical features of writing after the text has been produced; it does not check for a provider-specific watermark embedded during generation. The two approaches answer different questions and should not be treated as interchangeable. The National Institute of Standards and Technology’s overview of synthetic-content transparency approaches provides broader context on provenance, labeling, and detection.
Which ChatGPT writing is covered—and who can check it?
OpenAI’s October 5, 2026 announcement describes a limited rollout, not a watermark on every ChatGPT response:
- ChatGPT and Codex in the EU: eligible text output is to receive invisible watermarks over the coming weeks. OpenAI did not announce a global ChatGPT default at launch.
- API use: customers globally can opt in for select models. Watermarking is off by default in the API.
- Detector access: OpenAI is initially limiting its text detector to approved researchers and expert organizations. Access applications are reviewed case by case.
Therefore, you cannot assume a passage carries a watermark just because it may have come from ChatGPT, or that a public checker is available to test it. Details of the rollout and eligibility are in OpenAI’s EU text provenance announcement, Help Center explanation of provenance signals, and developer documentation on content provenance.
How well did textGrain perform in OpenAI’s tests?
OpenAI reported these results from separate evaluations. They are vendor-reported outcomes under the stated conditions, not guarantees for any arbitrary passage:
| Evaluation | OpenAI-reported result | What the condition means |
|---|---|---|
| Psychology passages, 200 tokens | About 80% detected | Tested at a target 1% false-positive rate; OpenAI evaluation, 2026. |
| Psychology passages, 400 tokens | About 95% detected | Tested at a target 1% false-positive rate; OpenAI evaluation, 2026. |
| Editing evaluation, 400-token passages | About 92% baseline detection | Detection fell to about 66% after replacing 10% of words with synonyms, and to 17% after replacing 25%; OpenAI evaluation, 2026. |
Detection was substantially lower for mathematics, where there are fewer natural choices of wording. The figures come from different tests and should not be combined into a single accuracy estimate. In particular, detection rate is not the same as a guarantee that a positive result is correct in every real-world setting.
What a positive or negative result means
A positive result is a signal, not an authorship verdict
A positive check means the detector found a supported OpenAI watermark signal. OpenAI says, “A watermark does not identify the user.” It does not reveal an account, prompt, or conversation, nor does it measure how much a person edited or contributed. It cannot establish ownership, responsibility, legality, or factual accuracy.
A negative result does not rule out ChatGPT use
The text may come from a legacy model, predate supported watermarking, or have been edited enough to weaken the pattern. A passage may also be too short or constrained for the detector to identify the signal reliably. OpenAI’s API documentation says its provenance checker does not currently detect content generated by another company’s AI system.
Can you ask ChatGPT whether it wrote a passage?
No. OpenAI warns that ChatGPT has no reliable knowledge of whether it generated a particular passage and may invent an answer when asked. A chatbot’s claim that it did—or did not—write text is not verification. See OpenAI’s guidance, “Can I ask ChatGPT if it wrote something?”
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.How watermark checks differ from AI-writing classifiers
| Method | What it checks | Main limitation |
|---|---|---|
| OpenAI text watermark detector | A provider-specific statistical signal embedded during eligible text generation. | Only applies to supported output where watermarking was enabled; access is restricted at launch, and edits can reduce detection. |
| Third-party AI-writing classifier | Patterns in wording that a classifier associates with AI-generated text. | It does not verify an OpenAI watermark or prove that ChatGPT produced the passage. |
| Asking ChatGPT | The model’s response to a question about authorship. | ChatGPT cannot reliably identify its own past writing, so the answer is not evidence. |
OpenAI’s discontinued AI Text Classifier illustrates why classifier results should not be confused with watermark detection. In an English challenge-set evaluation, that separate classifier identified 26% of AI-written examples as “likely AI-written” and mislabeled 9% of human-written text as AI-written. OpenAI discontinued it on July 20, 2023, citing low accuracy. Those figures describe the old classifier, not textGrain. See OpenAI’s announcement of the classifier and its discontinuation.
Quick Recap
How to interpret a claim that text came from ChatGPT
- Check whether the claim concerns an embedded watermark or a generic AI-writing classifier; they are different forms of evidence.
- For a watermark result, establish whether the text came from an eligible model and rollout, and whether watermarking was enabled.
- Consider passage length, subject matter, and editing history. Short, constrained, or substantially revised text can be harder to detect.
- Treat a positive result as evidence of a supported OpenAI signal, not proof of a person’s authorship or conduct.
- Treat a negative result as inconclusive about whether ChatGPT was involved.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
