The simplest way to think about it
When ChatGPT writes, it picks each next word from a long list of options. textGrain quietly tilts those picks, a little, toward a secret pattern. No single word looks odd. But across a few hundred words, the leaning adds up to a fingerprint that someone with the secret key can spot.
That’s why you can’t see it, and why copying and pasting keeps it: the fingerprint is made of the words themselves.
Why a longer text is easier to spot
OpenAI’s own numbers: at a one per cent false alarm rate, its detector caught about 80% of 200-token passages and about 95% of 400-token passages. That’s roughly 150 words versus 300. Maths answers were much harder, because there’s little choice about which words to use.
Why changing words weakens it
Since the pattern lives in your word choices, swapping words breaks it up. OpenAI found that replacing 10% of the words with synonyms dropped detection from about 92% to 66%, and replacing 25% dropped it to about 17%.
Who can see it
Only approved researchers and expert organisations, for now. OpenAI isn’t releasing the detector publicly because it can miss real watermarks and flag things that don’t have one. So a free website claiming to read textGrain isn’t something to rely on, and NeverPrompted doesn’t claim to.
Where you’ll run into it
If you use ChatGPT or Codex from within the EU, the mark is being switched on for every plan over the coming weeks. If you build with OpenAI’s API, it stays off unless you choose to turn it on for supported models. Text you generated in the EU and carried elsewhere keeps its pattern, because the pattern travels with the words.
Most people will never see a result. The detector isn’t public, so the realistic situations are researchers studying the technology, and organisations that have been approved to use it. If someone shows you a verdict, it’s fair to ask who ran it and with what key.
What it can’t tell anyone
It doesn’t say who wrote the prompt, how much of the final text is yours, or whether the text is true. And not finding it doesn’t prove a person wrote the text, which is OpenAI’s point as much as anyone’s.
Wherever this page describes a result: a detected mark is not proof of authorship, and an absent mark is not proof of human authorship. NeverPrompted's on-device rewrite can reduce detectable evidence but cannot guarantee defeating a vendor's undisclosed watermark, on any tier.