AnswersPublished19 sources
What OpenAI’s textGrain Watermark Means for ChatGPT and Codex
OpenAI’s textGrain subtly changes a model’s word choices to add a detectable statistical signal.
Published byEdited with GPT-6 LunaImages generated with GPT Image 2
Research answer
Key takeaways
- OpenAI’s textGrain subtly changes a model’s word choices to add a detectable statistical signal.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: What is OpenAI’s textGrain watermark, why is it introducing it for eligible ChatGPT and Codex text produced in the EU, how does its invisibl. Article summary: OpenAI’s textGrain is an invisible statistical watermark: it subtly changes the pattern of words a model chooses so a detector can look for a signal that the text came from OpenAI. OpenAI is introducing it to help make A. Topic tags: general, general web, user generated. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, charts with fa
OpenAI’s textGrain is an invisible statistical watermark designed to help identify text generated by its models. Instead of adding a visible label, it subtly influences word choices so a detector can check for a corresponding signal. OpenAI plans to add it to eligible ChatGPT and Codex text generated in the EU, while offering an opt-in version for some API models worldwide. It can help establish provenance, but it is not a definitive test of who wrote a passage—or whether text without a detected watermark is human-written. 12
13
How textGrain works
TextGrain changes the randomness behind a model’s word choices during generation. The resulting pattern is statistical rather than a hidden character or visible mark in the text. A detector checks for that pattern using the corresponding key; it does not determine whether writing simply sounds AI-generated. 5
13
That distinction matters: a detection indicates that a text contains a watermark signal associated with OpenAI, not who prompted the model, how much a person edited the result, or who published it. A missing signal is not proof that a person wrote the text, because the watermark may be difficult to detect or may have been weakened. 6
29
Who gets the watermark—and who can detect it?
OpenAI announced the rollout on October 5, 2026. It said eligible ChatGPT and Codex text generated in the EU would receive textGrain over the following weeks, across plans. This is a phased rollout, not a statement that every eligible output was marked immediately on announcement day. 12
10
API customers worldwide can opt in to watermarking for select models. The feature remains off by default in the API, and OpenAI said it would not become a global default at launch. 12
11
Watermarking access and detector access are separate. OpenAI opened applications for its detector, initially limiting access to approved researchers and expert organizations that can help evaluate the technology. It has not announced an open public detector. 12
Why OpenAI is introducing it in the EU
The rollout responds to the EU AI Act’s requirement for generative AI providers to make generated text identifiable in a machine-readable way. Reporting on the rules says the relevant transparency provisions took effect on August 2, 2026; OpenAI’s announcement described adding the watermark over the weeks that followed. 12
10
Machine-readable provenance gives other systems a way to check for a signal that may not be visible to readers. But textGrain is one company’s watermark, not a universal label or guarantee that every AI-generated passage can be recognized. OpenAI says the system performed as well as or better than approaches it evaluated, including Google’s SynthID for text. That is OpenAI’s reported comparison; it does not mean the systems or their detectors are interchangeable. 3
9
What can make textGrain detection fail?
Detection is less dependable when there is little text to analyze or when the wording has changed substantially. OpenAI’s help information says extensive paraphrasing or translation can make a watermark undetectable, even though the system is designed to withstand some common changes. 13
One report of OpenAI’s testing says that replacing a quarter of the words reduced detection to 17%. That result is specific to the reported test; it should not be treated as a universal detection rate for all text, models or edits. 6
The practical takeaway is to treat a positive result as evidence of a watermark signal—not proof of a person’s role or intent. Treat a negative result cautiously, too: edited, translated or short text may carry a signal that the detector cannot reliably find.
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "What OpenAI’s textGrain Watermark Means for ChatGPT and Codex"?
OpenAI’s textGrain subtly changes a model’s word choices to add a detectable statistical signal.
Sources
openai.comOur approach to EU text provenance rules
help.openai.comProvenance signals (Content Credentials, SynthID) in OpenAI ...
fonearena.comOpenAI textGrain watermarking for ChatGPT and Codex to roll out in EU
xenospectrum.comOpenAI to Add Invisible Watermarks to ChatGPT Text in the EU, With About 95% Detection Under Set Conditions