OpenAI Unveils “textGrain”: Invisible Watermarks Are Coming to ChatGPT Text

October 5, 2026

OpenAI has announced how it will comply with the EU AI Act’s requirement that AI-generated text be machine-identifiable: an invisible watermarking technology it calls textGrain. API customers can opt in starting today, and in the coming weeks the watermark becomes automatic on eligible ChatGPT and Codex output in the European Union — with no off switch. It’s a significant step toward making “made by AI” labels a standard part of text generation.

What happened

OpenAI announced textGrain on Monday, its approach to the EU AI Act’s transparency obligations for generative AI systems. The rollout is deliberately cautious: API customers globally can opt in to watermarking for select models starting today, while in ChatGPT and Codex the invisible watermark will be added automatically only to eligible output in the EU, in the coming weeks. Access to OpenAI’s watermark detector isn’t public at launch — applications are open, but initially limited to approved researchers and expert organizations while the company evaluates and improves the technology.

How textGrain works — and its limits

textGrain adds an invisible statistical signal to the model’s word choices, nudging which synonyms and phrasings get selected in a pattern a detector can recognize later. OpenAI published a technical report on the method and says it plans to open-source the technology so others can build on it. Intriguingly, OpenAI’s own comparison showed watermarked outputs scoring slightly higher than unwatermarked text on several benchmarks, including its internal Astra-max evaluations.

But OpenAI is unusually candid about what the watermark can’t do. Detection can produce both false negatives and false positives, particularly with short texts or domains like mathematics where word choice is constrained. And it’s fragile: in OpenAI’s testing on 400-token passages, replacing just 10% of words with synonyms cut detection from about 92% to 66%; swapping 25% collapsed it to 17%. The watermark also “does not measure human contribution,” doesn’t establish ownership or responsibility, doesn’t identify the user, and doesn’t verify accuracy. Absence of a detected watermark, the company stresses, does not prove human authorship.

Why it matters

Anyone hoping for a reliable AI-detector story should temper expectations — light paraphrasing defeats the signal, and OpenAI itself says a clean scan proves nothing. That’s the honest truth, and it’s good OpenAI says it up front. The bigger story is that watermarking is going from policy to plumbing: Anthropic added a similar word-choice watermark to Claude back in August (not without controversy), Google DeepMind has been watermarking outputs for a couple of years, and now OpenAI is joining in — with an open-source release planned. For everyday users, the near-term impact is minimal: the mandatory rollout is EU-only and limited to ChatGPT and Codex for now. But the direction is clear: within a year or two, “this text was generated by an AI” may be a standard, machine-checkable property of the text itself — at least when nobody has bothered to paraphrase it away.

FAQ

Is the watermark visible?

No. It’s an invisible statistical pattern in word choices — readers won’t notice anything different about the text.

Will ChatGPT users outside the EU get watermarked?

For now, mandatory watermarking applies only to eligible ChatGPT and Codex output in the EU. API users anywhere can opt in voluntarily starting today.

Can people just paraphrase to remove it?

Largely, yes. OpenAI’s own tests show swapping 25% of words drops detection to 17%, so light editing defeats it.

Does a detected watermark prove text is AI-written?

No. OpenAI acknowledges false positives and false negatives, and says the watermark doesn’t measure human contribution or establish authorship.

Why is this happening now?

The EU AI Act requires providers of generative AI systems to mark AI-generated text in a way that makes it machine-identifiable, which pushed the labs to ship watermarking systems.

Sources: OpenAI, 9to5Mac, Gizmodo