Does the ChatGPT watermark show up in client messages?

What is ChatGPT's new watermark, and who gets it first?

On October 5, 2026, OpenAI announced an invisible, machine-readable watermark for text written by ChatGPT and Codex. The system is called textGrain, and for now it only applies to users in the European Union. It will reach eligible EU users on every plan, free and paid, over the coming weeks. OpenAI hasn't announced the same for ChatGPT users outside the EU. Developers anywhere in the world who build on OpenAI's API can opt in for selected models, but that setting is off by default. So unless you're in the EU, or a developer has deliberately switched it on, the ChatGPT text you copy today carries no watermark at all.

The timing comes from the law, not from a sudden change of heart. Article 50 of the EU AI Act sets transparency rules that took effect on August 2, 2026. They require providers of generative AI to mark their output in a way other systems can detect. OpenAI actually built a text watermark back in 2024 and then shelved it. It worried about false accusations against people writing in a second language, and about users moving to competitors. In August, Anthropic said it would watermark Claude's text using a version of the SynthID approach that Google DeepMind published in 2024. OpenAI now says textGrain matched or beat approaches like SynthID in its own internal tests.

So why should a plumber in Lisbon or a consultant in Berlin care? Because a lot of people who run small businesses now draft client messages in ChatGPT and paste them straight into WhatsApp. Think payment reminders, quote follow-ups, "just checking in" nudges and polite chases for signed documents. The obvious question is: can my client tell? That's a fair thing to ask, and it deserves a straight answer rather than panic or a shrug. The rest of this article covers what the watermark really does, what it can't do, and the habits that keep your AI-assisted messages sounding like a person your clients already know.

How does textGrain hide a watermark in ordinary text?

textGrain hides a pattern in the words ChatGPT chooses. It doesn't add hidden characters or tags. When the model writes, there are often several words that would fit equally well: "quick" or "brief", "send" or "share", "soon" or "shortly". A secret key quietly decides which of those equal options gets picked. One word proves nothing, because any human could choose "shortly". Over a long enough passage, though, those small nudges add up to a statistical fingerprint. Someone holding the key can test for it. Everyone else just sees normal sentences, and the writing doesn't look or read any differently.

This matters if you copy and paste. A common worry about AI watermarks is invisible characters, like zero-width spaces, riding along into your message. OpenAI says textGrain doesn't insert anything like that. What you paste into WhatsApp is plain text, letter for letter what you see. It doesn't carry metadata about who you are, what plan you're on or what you typed into the prompt. The detector only reports whether it finds an OpenAI watermark. It doesn't track who wrote what, and it doesn't check whether the content is accurate. OpenAI also reported practically identical benchmark scores with and without the watermark, so you shouldn't expect your drafts to get any worse.

The limits come straight from OpenAI's own figures. Under ideal conditions, the detector catches about 80% of OpenAI text at 200 tokens and about 95% at 400 tokens. If a quarter of the words are swapped for synonyms, detection drops to 17%. OpenAI's tests also show that paraphrasing, translation, short passages, code and formal text can weaken or wipe out the signal. A token is roughly three-quarters of an English word, so 200 tokens is around 150 words. Now picture a typical client follow-up on WhatsApp: "Hi Dana, just checking whether you've had a chance to sign the quote. Happy to jump on a call if anything's unclear." That's about 25 words.

Can clients detect ChatGPT text in your WhatsApp messages?

Clients can't realistically spot ChatGPT text in your WhatsApp messages through the watermark. There are two reasons. Your messages are far too short to carry a reliable signal, and your clients don't have the detector. Right now only approved researchers can use it, and they have to apply for access. OpenAI plans to open-source textGrain eventually, which could put detection into more hands. Even then the length problem doesn't go away. A two-line nudge sits well below the few hundred words the system needs to make a confident call. Detecting a watermark in a WhatsApp chat about an unpaid invoice is like trying to identify a song from one note.

Most real-world messages also go through steps that weaken the signal even more. Plenty of business owners in Israel, Latin America and across Europe draft in English and then translate, or ask for a draft in one language and send it in another. Translation is one of the things OpenAI's own tests flag as damaging the watermark. So is light editing: change a greeting, cut a sentence, add a name, and the pattern gets thinner. Then there are the online "AI detectors" some people paste text into. Those tools don't read watermarks. They guess from writing style, they're known to wrongly flag human writing, and very few clients are going to run your message about Thursday's delivery through one.

Here's the part people overlook: the watermark was never the real risk. Your voice is. Clients don't need a detector when a message opens with "I hope this message finds you well" and closes with "Please don't hesitate to reach out." They notice bullet points in a chat thread, a three-paragraph reply to a yes-or-no question, or a sudden formal tone from someone who usually writes "hey, any news?" People know how you text. If your messages suddenly sound like a customer service template, they'll pick up on it faster than any software could. Worrying about invisible watermarks while sending messages that obviously sound like AI is putting your energy into the wrong problem.

Should you tell clients when AI helps write your messages?

You don't need to label every message you tidied up with AI, but you should never let an AI pretend to be you in a conversation your client thinks they're having with you. That's the line that matters. Using ChatGPT to polish a draft is like using spellcheck or asking a colleague to read something over: the words go out under your name, and you've approved them. An AI that holds a whole conversation while posing as you is a different thing, and clients are right to feel misled when they find out. Transparency is also the clear direction of the new EU rules. People should know when they're dealing with a machine.

When AI drafts and you hit send, the message is yours, and so is responsibility for it. Read every draft before it goes out, and check anything with a number in it. AI tools are good at sounding sure about details they've made up: a due date that was never agreed, the wrong invoice total, a discount you never offered. Check the client's name, too. It's an easy mistake when you're working through ten follow-ups in a row. A quick read takes fifteen seconds. Explaining why your message promised delivery "by Friday" when you meant next month takes a lot longer, and costs you more goodwill than any watermark ever could.

There's a more honest option when you'd rather not do the chasing yourself: have an assistant that openly says it's your assistant. That's how outreach tasks work in Agent Sonic. You tell it in plain words, "Ask Dana when the signed quote will be back and let me know", and it shows you the exact first message before anything is sent. Sonic writes to Dana from its own WhatsApp number, introduces itself as your assistant and never pretends to be you. It follows up at most once if the answer is vague, then reports back. It's available on the Pro plan, it's in Beta and may be paused from time to time, and the step-by-step tutorial on having an assistant message people for you walks through the limits.

How do you write AI client follow-ups that still sound like you?

Give the AI a short, specific brief, then edit the draft instead of pasting it raw. That's the whole habit. A vague prompt like "write a follow-up to a client" gets you a generic letter that could come from any company. A tight one gets you something close to how you'd actually write it: "Chase Dana about the signed quote, friendly, two lines, no emoji." Say how long it should be, what tone you want, what language to use and the one thing you want back. Then treat the result as a first draft from a capable junior who doesn't know your client yet. They get you most of the way there, and you finish it off.

A few edits make the biggest difference. Keep your usual greeting and sign-off, whether that's "Hi Dana" or just "Dana,". Cut the filler opener, because nobody you text weekly needs to hear that you hope they're well. Stick to one request per message. Add one detail the AI couldn't know, like "hope the move went OK" or "great to see you at the market on Saturday". Then read it aloud: if it sounds like something you'd never say out loud, rewrite it. You can also set your style once. Tell an assistant with memory "I never use emoji with clients, keep it under 40 words, sign off as Sagi" and it applies that every time.

You'll save the most time by doing all of this without leaving WhatsApp. Agent Sonic works like ChatGPT that lives in your WhatsApp: you message it like any contact, ask for a draft, tweak it in the same chat and copy it into the client's thread. Then add the step most people forget: "Remind me Thursday at 10 to check whether Dana signed." The follow-up stops depending on your memory. There's nothing to install and no dashboard to check. If you want your admin handled in the app you already have open all day, get access to Agent Sonic: leave your details and you'll get Sonic's number and a quick onboarding.

Frequently asked questions