AI Watermarks Just Became Law-Driven: OpenAI's textGrain in the EU
October 6, 2026 · 5 min read
Here is what textGrain actually is. When the model generates text, it has to pick each word from a huge list of candidates. A statistical watermark nudges that choice: it biases the model toward certain words in a subtle, patterned way that survives in the final output. You cannot see it. The sentence reads normally. But a detector, run later, can spot the pattern and say: this text was very likely machine-generated.
On October 5, OpenAI announced it will embed textGrain in text outputs from eligible ChatGPT and Codex users in the EU over the coming weeks, to meet the transparency requirements of the EU AI Act. Global API customers can opt in voluntarily, and researchers and expert organizations can apply for access to the detector. OpenAI says it has no plans to make it the default worldwide — for now. (via 新浪科技)
From "nice to have" to a legal requirement
That last line is the whole story. For years, watermarking AI text was an industry promise — labs said they were working on it, researchers published papers, and nothing shipped at scale. The EU AI Act changed the incentive structure. Its transparency provisions are not a suggestion; they are hard rules, and OpenAI is now complying with them the way a company complies with a tax code, not a pledge. This is the first brick: once one frontier lab embeds watermarks by law, the others operating in Europe face the same legal pressure to follow.
What a watermark can actually do
Its real job is scale, not courtrooms. A statistical watermark can flag large batches of machine-generated text — spam farms, SEO content mills, fake review operations, bot-generated propaganda. Platforms and publishers can run detection over suspicious content and quietly filter it out. That is a meaningful win: most of the damage from AI-generated slop is volume-based, and watermarks are a volume-based defense. Think of it as a metal detector at the door — imperfect, but it catches the obvious cases. Newsrooms can screen submitted op-eds, platforms can triage suspicious accounts faster, and regulators finally have a technical signal to point at when they talk about "machine-readable transparency."
What it cannot do
It cannot prove that any single article was 100% written by AI. Statistical signals are probabilistic, and they degrade. Rewrite the text, translate it to another language, or even paraphrase it aggressively, and the watermark weakens or disappears. Mixed human-machine editing makes detection fuzzy. So anyone promising a definitive "AI or not" verdict on a specific text is overselling. There is also a confidence problem: a detector that says "probably AI" will be wrong sometimes, and a single false accusation — against a student, a journalist, an author — does real damage. And there is a flip side: if detectors are widely available, attackers will learn to strip the signal — the cat-and-mouse game starts the moment the mouse knows what the trap looks like.
What you will notice: nothing
If you are an eligible EU user of ChatGPT or Codex, the practical impact is close to zero. The watermark lives in word-choice probabilities; it does not change how answers look, read, or behave. You will never see it. It only exists when someone runs the detector. For the API opt-in, the calculation is different: watermarked output is theoretically easier to identify, which is a privacy and competitive consideration each customer will have to weigh for themselves. And "not global default" is doing quiet work here — OpenAI is drawing a line between legal compliance in the EU and the rest of the world's expectations. That line may not hold for long. Once one jurisdiction makes watermarking normal, every other regulator will ask why their citizens don't get the same transparency. The question is not whether this spreads; it is who gets to set the terms.