Free tools Windows power users keep installed
One-click scans. No signup required.
Only a detector built to check OpenAI’s specific watermark can look for that signal. OpenAI says its new textGrain detector is initially available only to approved researchers and expert organizations. Generic AI detectors such as authorship classifiers do not verify a ChatGPT watermark, and even textGrain can miss watermarked writing—especially after editing or in subjects with less flexibility in word choice.
What does a ChatGPT watermark detector actually check?
OpenAI calls its system textGrain. It adds an invisible statistical signal through a model’s word choices, and a compatible detector checks for that signal. It is a provider-specific provenance check, not a general test of whether writing sounds like AI. OpenAI’s October 5, 2026 announcement describes the system and its initial rollout.
As an Amazon Associate I earn from qualifying purchases.
A generic AI text detector instead estimates whether text resembles examples of AI- or human-written prose. That estimate is not evidence that the text contains an OpenAI watermark. OpenAI’s text provenance documentation says its checker does not currently detect content from other AI providers.
Who can use OpenAI’s detector, and where is watermarking enabled?
In its October 5, 2026 announcement, OpenAI said API customers globally can opt in to text watermarking for select models; it is off by default in the API. The company said it would begin introducing invisible watermarking for eligible ChatGPT and Codex text output in the EU over the following weeks, across plans. It was not a global ChatGPT default at launch. Detector applications were initially limited to approved researchers and expert organizations. Those rollout details are specific to the announcement date; OpenAI said further technical details and open-source publication were planned, not already available.
How reliable is textGrain according to OpenAI?
OpenAI reported results at a target false-positive rate of 1%. In its evaluations, detection varied with both passage length and subject:
| Content and passage length | OpenAI-reported watermark detection |
|---|---|
| Psychology, 200 tokens | About 80% |
| Psychology, 400 tokens | About 95% |
| Mathematics | Substantially lower; a specific percentage is not stated in OpenAI’s announcement |
These are OpenAI’s stated evaluation results, not guaranteed accuracy across languages, topics, models, or real-world use. The lower result for mathematics is consistent with the challenge of detecting a word-choice signal where there is less freedom to vary wording. OpenAI’s announcement provides the reported figures.
Editing can weaken the signal
OpenAI also reported results for synonym replacement in 400-token passages: detection fell from about 92% without the described replacement to 66% when 10% of words were replaced, and to 17% when 25% were replaced. These are reported evaluation outcomes, not universal rates for every kind of paraphrase or editing. OpenAI’s announcement describes this test.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Scan for outdated or missing drivers - takes under a minute3Clear out junk files and repair common Windows errorsDoes a negative result prove ChatGPT did not write the text?
No. A “not detected” result does not establish that text is human-written or was never generated by OpenAI. OpenAI lists several reasons its checker might not identify OpenAI-origin material: stripped metadata, tampering, degraded watermarks, legacy models, or text produced before provenance signals were available. The checker also does not identify other companies’ AI output. See OpenAI’s text provenance guide.
A positive result also needs careful interpretation. It indicates a detected signal under the checker’s conditions; it is not, by itself, a complete account of who wrote, edited, or submitted a passage. OpenAI has described text watermarking and detection as early technologies with significant limitations. For consequential decisions, treat a detector as one limited signal and consider corroborating context such as drafts, version history, process documentation, and a fair conversation with the writer.
Can Turnitin, GPTZero, or another generic detector read the watermark?
The available information does not establish that Turnitin, GPTZero, or other named commercial detectors can read textGrain. Do not assume that a tool marketed as an AI-writing detector checks OpenAI’s watermark. Before relying on any product, verify whether it checks a provider-specific embedded signal or estimates authorship from writing features, which providers and models it supports, its language and length limits, and its false-positive and false-negative rates at a stated threshold.
Rank #4
Why generic AI detector scores are not watermark checks
OpenAI’s discontinued 2023 AI Text Classifier was an authorship classifier, not a watermark verifier. On its English challenge set, it correctly labeled 26% of AI-written text as “likely AI-written” and falsely labeled 9% of human-written text as AI-written. OpenAI said it was unreliable on short text, performed significantly worse outside English and on code, and should not be used as a primary decision-making tool. Those historical results apply to that classifier and test set—not to textGrain. OpenAI’s 2023 classifier announcement explains its limitations.
Other evidence reinforces the need to distinguish tools and conditions. A 2023 academic study found that recursive paraphrasing could significantly reduce detection rates for the detector types it evaluated; it was not a test of textGrain. The study addresses robustness to rewriting. Google’s SynthID Text is a separate system, not a ChatGPT watermark. Google says its detector is probabilistic and that thorough rewriting or translation can greatly reduce confidence; it also notes that watermarking is less effective when factual precision leaves little room to vary word choice. Google’s SynthID Text documentation describes that system.
How to use detector results responsibly
- Identify what is being tested: an OpenAI watermark check and a generic AI-authorship estimate answer different questions.
- Check scope: confirm the relevant provider, model, language, passage length, and subject are supported.
- Read the threshold and error rates: a detector score is meaningful only in context, including how often it misses signals and how often it flags text without them.
- Account for editing: paraphrasing, translation, or other changes may weaken a watermark or shift a classifier’s estimate.
- Do not make a high-stakes judgment from one score: use process evidence and a fair review alongside any detector result.
The European Commission’s technical report groups text provenance approaches into watermarking, structural marking, metadata, logging, and AI-generated-text detection. It suggests comparing them by effectiveness, robustness to editing, reliability across scenarios, user accessibility and interpretability, and interoperability. That framework helps explain why “reliable” is not a single property shared by all detectors. Read the European Commission’s report.
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.
