OpenAI is preparing to embed an invisible watermark in text written by ChatGPT and its Codex models, so that AI-written messages, blog posts, essays, emails and research papers can be identified afterwards. The company says the feature will launch first for users in the European Union, to comply with the bloc's AI Act. The technology is called TextGrain, and because it is still in development the rollout may slip by several weeks.
How you watermark something made of words
Watermarking an image is intuitive: you alter pixels a viewer will not notice. Text has no pixels, so the technique has to work somewhere else — in the choices the model makes.
At every step a language model is picking the next word from a ranked list of plausible candidates. Usually several are about equally good. A watermark exploits that slack: the model is nudged to prefer one subset of acceptable words over another, according to a secret pattern. Any single sentence looks completely normal, because every word chosen was a word the model might have chosen anyway. Across a few hundred words, though, the distribution of those choices carries a statistical signal that software holding the key can measure.
That is why OpenAI describes it as a signal in the words and sentences rather than a tag attached to the file. There is nothing to strip out, because there is nothing bolted on.
What removes it anyway
The method is elegant and it has a plain weakness: it survives copying and pasting, and it does not survive rewriting.
Change enough of the words and the statistical pattern washes out, because the pattern is the words. Paraphrasing by hand weakens it. Running the text through a different AI model to reword it weakens it a great deal, since the second model imposes its own word choices and knows nothing of the first one's key. Translation almost certainly destroys it.
Length matters too. The signal is statistical, so it needs enough text to rise above chance — a long essay can be checked with confidence, a two-line email cannot.
So this is not a lie detector for homework. It is a reasonably strong signal on long, unedited, machine-written text, and close to useless on anything a motivated person has passed through a second tool.
Why the EU first
The sequencing answers the obvious question about motive. OpenAI has had the research capability to do this for years and did not ship it. The EU AI Act requires providers to mark AI-generated content in machine-readable form, and the feature is appearing where that requirement bites.
Which tells you what to expect elsewhere. Markets with a comparable rule will get it; markets without one may not, or may get it later and quietly. Bangladesh has no such rule, so whether a Bangladeshi user's ChatGPT output carries the mark will be decided by OpenAI's engineering convenience rather than by any local policy.
What it changes for a student or a freelancer here
Very little in the short term, and the honest advice is not about evasion.
A detector that only the issuing company can run is not something a Dhaka university or an international client will be querying. Institutions will keep using third-party detectors, which work by guessing rather than by reading a key, and which are wrong often enough to ruin someone's week — a point already live in the row over Turnitin and in a literary prize dropping a novel over an accusation.
What does change is the direction of travel. Marking machine output is becoming a legal requirement rather than a voluntary courtesy, and the platforms are building the plumbing. The parallel effort on images and video is the C2PA provenance standard now live in TikTok — which shares this one's shape, including that a screenshot defeats it.
On teaching people to read AI output critically rather than relying on detection at all, see the school that hired a teacher for it.
