By the Smili Phone team
Starting this month, some ChatGPT answers will carry a secret signature you can never see. OpenAI announced on October 5, 2026 that it will embed an invisible watermark called textGrain into text and code produced by ChatGPT and Codex for users in the European Union, rolling out over the coming weeks. The reason is a new law: the EU AI Act requires AI-generated text to carry a machine-readable label, and existing providers have until December 2 to comply. Readers cannot see it, but special detection software can test a passage for the signature.
How the invisible ink works
textGrain does not hide characters in the text or use sneaky spacing tricks. Instead, it gently nudges the model's word choices as it writes, following a secret key, creating a statistical pattern woven into the language itself, like a hallmark pressed into silver. Because the mark lives in the word choices, it survives a simple copy and paste, and stripping hidden characters will not remove it. OpenAI says the watermark does not identify the user, reveals no prompts or conversations, and caused no meaningful change in writing quality, and the company plans to open source it.
Why Europe, and why now
The EU AI Act's transparency rule says providers of generative AI must make their text outputs identifiable in a machine-readable way. Anthropic made a similar move in August, watermarking all Claude output worldwide with no opt-out, and Google's SynthID system has worked this way for years. The United States has no equivalent federal rule, so OpenAI is keeping the watermark mandatory only inside the EU. Developers anywhere can switch it on voluntarily from October 5, but it stays off by default.
What the watermark cannot do
This part matters. The detector is good, not magic. OpenAI's own tests found it catches the mark in about 80 percent of 200-token passages and about 95 percent of 400-token ones. Editing the text weakens it fast: swapping one word in ten for a synonym dropped detection from roughly 92 percent to 66 percent, and swapping one word in four cut it to about 17 percent. Short texts, math, and translated text are all harder to verify. For now, only approved researchers and expert organizations get the detector; it will not be a public feature. And OpenAI warns that a missing watermark does not prove a human wrote something. It cannot tell you who wrote a passage or whether the information in it is true.
What this means for you
Most of us will never touch the detector, but the idea behind it is worth understanding: AI-written text is everywhere now, and telling it apart from human writing is getting harder. Scam messages are a good example. Fraudsters already use AI to write convincing fake bank alerts and delivery texts, the kind we wrote about in our guide to mobile banking scams. A watermark system like textGrain is one more tool for catching that kind of fakery, even starting in Europe.
Until detection goes mainstream, your own judgment is still the best filter. Be suspicious of any message that pressures you to act fast, and do not tap links in texts you did not expect. If a message on your phone looks strange and you are not sure it is real, bring the phone to our shop at 1882 W El Norte Pkwy in Escondido and we will look at it with you. No appointment needed.
Related: fake iPhone Duo pre-orders spread malware · your phone's hidden security mode
