What changed
OpenAI announced October 5 that ChatGPT and Codex will add an invisible watermark to text generated for users in the European Union, rolling out over the coming weeks. The watermark, called textGrain, works by nudging the model's word choices to create a statistical pattern a detector can spot. API users anywhere can opt in starting October 5; it stays off by default outside the EU.
The move answers Article 50 of the EU AI Act, which requires AI-generated content to be machine-readable. Only approved researchers and organizations get access to the detector for now.
What you can do with it now
Knowing the watermark exists—and where it breaks—lets you use AI drafts with more confidence. If you're in the EU and need to show a client or regulator that a document came from your team, not a chatbot, light human editing already makes detection unreliable. OpenAI's own tests show that swapping 10 percent of words for synonyms drops detection from 92 percent to 66 percent. Swap 25 percent and it falls to 17 percent.
If you hand ChatGPT a task and then rewrite the result in your own words—the workflow most small teams already use—the watermark mostly disappears. That's useful information: it means the tool won't flag work you actually shaped.
What to keep in your own hands
Don't treat watermark detection as proof of authorship. OpenAI states clearly that the watermark cannot tell you who wrote a passage, how much a person edited it, or whether the content is accurate. A missing watermark proves nothing about human authorship either—text could be too short, heavily edited, or from another AI tool.
If a client, partner, or compliance process asks you to verify your work isn't AI-generated, the watermark won't solve that. Keep records of your drafts, edits, and decisions. The work you can defend is the work you can show you shaped, not the work a detector didn't flag.