Last week OpenAI announced that ChatGPT will start hiding a signature inside the text it writes. If you use it to answer guests, draft quotes or tidy up an email, you may be wondering whether your customers will now see an "AI wrote this" stamp on your messages.
They won't. But what the signature is, and what it can and can't prove, is worth five minutes of your time.
What a text watermark actually is
It is not a label. It is a habit.
When a model writes, it picks each next word from a short list of likely candidates. A watermark uses a secret key to nudge which candidate wins. No single sentence looks odd. Over a few hundred words, though, the choices form a pattern, a bit like a person who always picks the second-most-obvious synonym. Someone who knows the habit can spot it on a full page. Nobody else can.
According to TechCrunch, OpenAI calls its version textGrain. It is rolling out over the coming weeks, automatically for people in the EU on every ChatGPT and Codex plan. For developers using the API it is optional and off by default.
Why it is happening now
Article 50(2) of the EU AI Act says that companies providing AI systems that generate text must mark the output in a machine-readable way. It applies since 2 August 2026. Systems that were already on the market, ChatGPT among them, have until 2 December 2026 to catch up. That is my reading of why the rollout is happening in these weeks.
The key point for you: the duty sits on the company that builds the model, not on the person typing the prompt. Anthropic has described a similar mark for Claude, according to ACS Information Age, and Google already uses its own system, SynthID.
What you have to do about it
For the watermark itself, nothing.
There is a separate duty for people who use AI (Article 50(4)), but it covers text published to inform the public on matters of public interest, and even then it falls away when a person reviews the text and takes editorial responsibility. A quote, a review reply or a booking confirmation is not that. If you are unsure, my earlier piece on the EU law walks through who owes what. I'm not a lawyer, and sector rules can add more, so check yours.
One rule does reach small businesses: if a customer talks directly to an AI, say a WhatsApp assistant, they should be told. A watermark does not replace that sentence.
What it cannot do
The watermark is fragile, and OpenAI says so. Per TechCrunch, swapping 10% of the words for synonyms dropped detection from 92% to 66%. Short passages, math answers and translated text are harder to detect. Rewording removes it. Sean Goedecke explains why: text is so compressed that any change to a sentence is a change you can see, so the signal has nowhere to hide.
OpenAI's own line is the one to remember: "A missing watermark does not prove human authorship." Two traps follow from that.
The first is worry. A message you read and edited probably carries little of the signal anyway.
The second is trust in detectors. Say you run a holiday apartment and a cleaning supplier's email "has no watermark". That tells you nothing. And if a job applicant's letter "has one", that is a weak statistical hint, not proof of anything. Please never accuse an employee or a candidate on that basis.
Where I'd stay cautious
I could not find whether OpenAI will give the public a detector, so I would not assume customers can check your text at all. Other tools (Gemini, Claude, open models) do this differently or not yet. And the EU rule covers EU users, so the picture elsewhere is less clear.
One step this week
Write down the three places where AI-written text leaves your business under your name: replies, quotes, listings. Beside each one, note whether a person reads and changes it first.
For most owners the answer is "yes, mostly". That habit does more for your customers than any hidden signature, and it is also what stops your text from sounding like everyone else's. Where the answer is "no", that is the place to start.