Every paragraph ChatGPT writes for a European user will soon carry a signature you cannot see.
OpenAI announced on October 5 that it will start watermarking ChatGPT's text in the EU, TechCrunch reported. The rollout covers ChatGPT and Codex users across all plans in the coming weeks. Developers using the API anywhere in the world can switch it on now, though it is off by default.
The driver is regulation. The EU AI Act transparency obligations took effect on August 2, requiring providers to mark AI-generated content in a machine-detectable way. I covered the first wave of that when the AI label became law. Now it reaches plain text.
How textGrain actually works
OpenAI calls the method textGrain. It works by subtly shaping the model's word choices, leaving a statistical pattern that a reader cannot see but a detector can pick up. There is no hidden character to delete and no metadata to strip.
That design has an obvious weakness, and OpenAI published it. In its own testing, replacing 10 percent of the words with synonyms dropped detection from about 92 percent to 66 percent. Heavier editing pushes it lower still.
Access to the detector is also limited. For now it goes only to approved researchers and expert organisations, so your competitor cannot paste your blog post into a checker tomorrow. The technical report was co-authored with researchers from the University of Pennsylvania and Yale.
This follows Anthropic, which announced global watermarking for Claude in August. That announcement drew a backlash from users worried about workplace and academic integrity. Two of the largest model providers now sign their text. The third and fourth will not hold out for long inside the EU.
Weak watermark, strong signal
The easy reaction is to shrug. If a light edit beats it, why care? Because the watermark was never designed to catch a determined editor. It was designed to make raw, unedited machine output identifiable at scale.
That is exactly the content that causes problems. The product page generated in bulk and never reviewed. The cold email sequence pasted straight from the chat window. The thought leadership post that nobody at the company actually read before it went live.
A watermark that survives only when nobody touched the text is, in practice, a detector for lazy workflows. That is the part content teams should think about.
There is also a second-order effect. Once detection exists, even restricted, the question changes from "did you use AI?" to "can you show what you did with it?" Editing, fact-checking and adding original material stop being quality habits. They become the evidence trail.
What changes for marketing and content teams
For EU-facing businesses, the practical question is disclosure. The AI Act already pushes deployers toward labelling certain AI-generated content. A watermark in the underlying text makes a silent policy riskier, because the fact of AI assistance may be verifiable later by someone else.
I would rather decide that policy myself than have it decided by a detector result in a client dispute. At difrnt., the rule I like is simple: AI can draft, humans own the claim. Every factual statement, every number and every promise in published copy has a named person who checked it.
Agencies and freelancers need to read their contracts too. Many service agreements promise "original work". A detectable machine pattern in the final file opens a conversation that is better had at the start of the engagement than at renewal.
And if you build products on the API, check your settings. Watermarking is off by default for developers, but EU customers may start asking whether outputs are marked. That is a product decision, not a toggle to discover in a sales call.
The authenticity premium gets more expensive
There is a commercial angle beyond compliance. When two provenance stacks landed in the same week earlier this year, the bet was that verified origin would become a quality signal. Text watermarking pushes that bet into everyday copy.
If machine text becomes identifiable, unedited machine text becomes a visible cost-cutting signal. Readers already discount content that sounds generic. Platforms are already reducing the reach of low-effort material. A traceable pattern gives both of them a cleaner way to sort.
The answer is not to hide AI use. It is to put things into the work that no model could have produced on its own: first-party data, named sources, a specific client situation described without breaking confidentiality, a judgement call someone was willing to sign.
What I would do before the rollout lands
Write a one-page AI content policy this month. State what AI is used for, who reviews the output, and when you disclose it. Keep it short enough that people actually follow it.
Audit the content you publish at volume, the product descriptions, programmatic pages and email sequences. Those are the places where unreviewed text is most likely to sit, and where a watermark is most likely to survive intact.
Then add a review step that changes the text for a real reason: correcting facts, adding data, tightening the argument. That makes the content better, and the evidence trail follows on its own.
The watermark will not catch good work. It was built to catch work nobody bothered to finish.