No. If OpenAI’s textGrain detector does not find a watermark, that does not prove a person wrote the passage. The text may be too short or constrained, edited or translated, generated by an unsupported model, or produced before watermarking was enabled. A positive result is limited too: it suggests OpenAI model involvement, but does not establish who wrote the text, how much a person contributed, or whether the claims are accurate.
What is OpenAI textGrain?
OpenAI describes textGrain as a text-watermarking technology that embeds a statistical signal in a model’s word choices. During generation, it subtly adjusts how the model selects among possible next words or word pieces. The signal is not a visible mark, hidden character, invisible space, unusual punctuation, or special watermark token.
As an Amazon Associate I earn from qualifying purchases.
OpenAI’s Help Center explanation describes the underlying method: at each generation step, the system uses a secret key to favor different token choices. The resulting distributions are balanced so their average matches the model’s original likelihoods. A detector with the matching key and settings checks whether the passage follows the expected pattern more often than chance.
Recommended Free Tools
This differs from a classifier-based AI detector. A classifier examines text after it is written and infers whether its patterns resemble AI output; textGrain looks for a signal inserted during generation. Neither kind of result, by itself, proves who authored a passage.
#1 Best Overall
What does a missing watermark mean?
It means only that the detector did not identify a watermark in the text it evaluated. OpenAI says detection can fail for several reasons:
- Short or constrained text: A short passage contains fewer choices from which to detect a statistical pattern. Precise answers, including some math, also leave less room for varied wording.
- Editing or translation: Replacing words or translating text can weaken the signal. OpenAI’s tests found that synonym replacement reduced detection substantially.
- Coverage and timing: The passage may come from a model or product that does not support watermarking, or it may predate the feature’s rollout.
- Another provider: Text generated by a different company’s system is not expected to carry an OpenAI watermark.
OpenAI says the EU Code of Practice does not require watermarks for outputs under 200 tokens—about 150 English words—or for code snippets. That threshold is not a promise that every longer passage can be detected.
How reliable is textGrain detection?
OpenAI’s published results show why a detector result needs context. For content such as psychology, OpenAI reported about 80% detection for 200-token passages and about 95% for 400-token passages at a target false-positive rate of 1%. These are OpenAI evaluation figures, not a guarantee for an individual passage or a universal performance rate. Math detection was substantially lower than psychology detection because precise answers offer fewer wording choices.
Free tools Windows power users keep installed
One-click scans. No signup required.
Editing reduced detection in OpenAI’s 400-token tests: replacing 10% of words with synonyms lowered detection from about 92% to 66%, while replacing 25% lowered it to 17%. These figures describe the tested passages and conditions; they should not be read as a prediction for every edited text.
Rank #3
Language also matters. In a specific experiment using 500 synthetic English prompts translated into 24 official EU languages, OpenAI’s Help Center reported detection at a 1% false-positive rate of 69.0% for Spanish and 42.2% for Romanian. OpenAI says it adjusted watermark strength for languages below 60%. Those measurements describe that experiment, not expected accuracy across all writing in those languages.
False positives are possible: a detector can report a watermark when none is present. False negatives are possible too: it can miss a watermark that is present. OpenAI also reports no meaningful performance differences on its Astra benchmark evaluations, but that is the company’s benchmark report, not independent validation.
Rank #4
What can a positive result establish?
A detected watermark supports a narrow conclusion: an OpenAI system likely generated or processed some of the text. It cannot determine whether the model produced all of it or only part, whether a person made substantial edits or contributions, or which user or organization prompted the system. It does not establish ownership, responsibility, legal disclosure obligations, accuracy, or whether the text is presented in context.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Scan for outdated or missing drivers - takes under a minute3Repair Windows errors before they cause bigger problemsWho can use the detector, and where is watermarking available?
In its October 5, 2026 announcement, OpenAI said API customers worldwide could opt in to text watermarking for select models; it was off by default in the API. OpenAI also said it would add watermarking to eligible ChatGPT and Codex text output in the EU over the coming weeks. This is an announced rollout, not evidence that every eligible account or output is already covered. Model availability is selected in account settings and may change.
OpenAI said text-detector access was initially limited to approved researchers and expert organizations through a case-by-case application. The detector reports whether it finds an OpenAI watermark; it does not reveal the user, prompt, or conversation. This access restriction applies to text provenance; OpenAI’s image and audio verification tools have separate availability.
How should you use a detector result?
- Do not treat “no watermark detected” as proof of human authorship.
- Do not treat a positive result as proof that an entire passage was AI-written or that a particular person used AI.
- Use the result only as one limited provenance signal, alongside appropriate context and other evidence.
- Do not use a detector result alone to decide whether writing is truthful, original, properly disclosed, or attributable to a specific person.
OpenAI discussed a previous text-watermarking approach in 2024 and noted that translation or broad rewriting could undermine it. That discussion predates the 2026 textGrain announcement and is historical context, not a current textGrain test result.
Quick Recap
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.
