On Monday OpenAI announced it would stamp an invisible mark into the text ChatGPT generates for anyone in the European Union. The rollout reaches every plan over the coming weeks. Everywhere else, the same model, writing the same answer to the same prompt, writes nothing. No mark. No detector. Just text.
The mark is called textGrain. It shifts the model's word choices into a pattern no human reader notices but a detector holding the right key can read. OpenAI's own tests catch the pattern 92 percent of the time. The fine print arrives in the next paragraph. Change ten percent of the words and detection drops by nearly thirty points. Change twenty percent and it drops by nearly seventy-five. Translate it, shorten it, and the number falls again. The mark is at full strength the instant it is born and loses a little of itself with every edit.
That number says what the announcement was careful not to. A watermark that cannot survive a fifth of its own editing was never built to tell a reader what they are looking at. It was built to tell a regulator that the company complied. The two jobs look alike from a distance and separate on contact. A mark meant for readers has to survive the journey text actually makes, through an editor's hands, through a translator, through one sentence lifted into an email and retyped. A mark meant for regulators has to be present at the moment of generation, on the server, inside the jurisdiction. textGrain is the second thing.
The geography is the confession. The EU AI Act requires the mark, so Europe gets it. No American law requires it, so America does not. Nothing about the sentence changes when you cross the Atlantic. The same model, the same prompt, the same answer. One copy reads "AI-generated" to a detector. The other reads as plain prose. What differs is the address the operator filed them under. A fact about a machine's output has become a fact about where the machine's operator was standing.
OpenAI is not hiding who the mark is for. The detector is not public. The company is handing it to a limited list of researchers and organizations, with a request-for-approval process for everyone else. The person who would actually want to check a paragraph, a teacher, a journalist, a reader with a suspicion, holds no key and has no practical way to ask for one. The mark's real audience sits in a compliance department, and the mark is written for their checklist.
Anthropic turned its watermark on everywhere in August. OpenAI turned it on only where the law made it. Two companies read the same obligation and answered differently. One treated the mark as a property of the text. The other treated it as a property of the jurisdiction. Both positions are defensible. Only one pretends the reader was ever the point.
The mark changes nothing a reader can perceive. The words read the same and mean the same. The only person for whom the text is different is the one holding the key, later, and what the key reveals is a serial number, not a verdict on whether the writing is true. A generated sentence can be accurate and a hand-typed one can lie, and the watermark cannot tell those apart. It says only where the sentence came from, which is the easiest question in the room and the one that settles the fine.
The film business is watching this from the next room, because the same apparatus is aimed at frames. The EU AI Act already requires synthetic video to carry a machine-readable mark, and the marks exist, SynthID and C2PA among them. Ars Technica put it plainly: they are relatively easy for anyone with basic know-how to get around. And they degrade against exactly the work a filmmaker does for a living. Crop the frame. Re-grade the color. Re-encode for delivery. Run it through a stabilizer or an upscaler. Each pass is a normal finishing decision and each pass weakens the signal. The person most likely to touch the pixels after they are generated is the person the mark was supposedly going to expose, and their ordinary workflow scrubs it without trying.
So the filmmaker generating a shot in Paris and posting it to a global audience is labeled, and the same shot generated in Los Angeles is not. Nothing about the frame differs. The label tracks where the GPU was standing, and it vanishes in the edit bay regardless. For two years the argument in film has been whether AI footage should be labeled. The honest answer arriving this week is that labeling was never going to be universal, because labeling was never about the audience. It was about the operator's exposure, and exposure ends at the border.
None of this makes the mark useless. It makes the mark honest. A filmmaker who wants the audience to know can write a caption in any country, at no cost, and many do. That disclosure survives cropping and translation and time, because it lives in the record of who made the thing and said so, not in the pixels. The machine's mark was built for a different reader, in a different room, holding a different key.
The mark was never for you.
Bruce Belafonte is an AI filmmaker at Light Owl. He discloses in captions, which no border reaches and no detector key is needed to read.