What changed
On October 5, OpenAI announced textGrain, an invisible watermark that will appear in all ChatGPT and Codex text for EU users within weeks. The watermark subtly shapes word choices so a detector can identify AI-generated text with 95% accuracy on passages of 400 words or more. It is mandatory in the EU under Article 50 of the AI Act, which took effect August 2, 2026. For API customers worldwide, the watermark is optional and off by default.
What you can do with it this week
If you use ChatGPT or the API to draft customer emails, proposals, or web copy for European clients, expect watermarked output soon. If you are outside the EU, log into your OpenAI account settings and check the Text Provenance toggle under Data Controls to decide whether to turn it on. If you want customers to know which text came from AI, switch it on. If you want flexibility, leave it off while you can.
What to keep in your own hands
The watermark is not foolproof. OpenAI says accuracy drops from 92% to 66% if someone swaps just 10% of words for synonyms. Do not rely on it to prove authorship in a legal dispute or to catch every instance of AI text. Treat it as a signal, not proof. More important: the detector is restricted to vetted researchers and regulators at launch, so you cannot check text yourself yet. If labeling AI content matters to your work or your reputation, add your own disclosure instead of depending on a hidden mark only experts can read.