Can clients spot AI written messages on WhatsApp?

What did the new study find about AI writing tells?

A new study found that the newest AI models still leave fingerprints in their writing, and that those fingerprints keep changing shape. Published at the start of October 2026, the research compared 10,000 articles written by people before ChatGPT existed with AI rewrites of the same pieces. The models worked from summaries, so they couldn't copy the original wording. The team flagged roughly 13,000 words and phrases that showed up at least twice as often in the AI versions. Anthropic's Claude Opus 5.5 got the headlines. Its single biggest tell is the word “dependable,” which appears 23 times more often than in human writing, and it says “this matters” 116 times more often than people do.

There's some good news in there too. Opus 5.5 has almost completely dropped the em dash, cutting its use by about 99% compared with its predecessor. It has also stopped leaning on the old “it's not X, it's Y” construction. Now it says something “is more than an X, it's a Y.” Other models have their own habits. OpenAI's Astra loves what the researchers call corrective framing, writing that a topic is “not simply X.” That pattern turned up more than 100 times more often than in human text. The lead researcher summed it up bluntly: the tells aren't decreasing. Developers scrub out the famous ones, and each new model quietly picks up new ones.

Why should a plumber, a bookkeeper or a wedding photographer care about a word-frequency study? Because AI now drafts more of the messages you send clients, and some are even sent by an assistant on your behalf. A payment reminder that reads like a press release doesn't just sound odd. It tells the client you didn't really write it, and that makes it a little easier to ignore. Here's our take: you don't need to hide that you use AI. You need messages that sound like someone who knows the client, the invoice and the date. The rest of this guide shows how to get there, one message at a time.

How can you tell if a message is AI generated?

The fastest way to tell if a message is AI generated is to look for wording nobody uses in a chat: stock phrases, too much politeness and a sign-off that belongs in an email. On WhatsApp, people write short, skip the greeting half the time and mention specifics. AI drafts do the opposite. They open with “I hope this message finds you well,” explain why they're writing, repeat the request and close with “Please don't hesitate to reach out if you have any questions.” One of those might slip past. Put three in one message and your client knows exactly what happened, even if they couldn't tell you how they knew.

Here's a practical checklist of AI writing tells to scan for before you hit send. Over-explaining, where a two-line question is padded with context the client already has. Over-the-top warmth, like “I truly appreciate your time and understanding.” Empty adjectives such as “dependable” or “valuable” that add nothing. Signposting phrases like “just to clarify” or “this matters because.” Lists that always have exactly three items. Perfect punctuation and full paragraphs in a medium where people send fragments. And the biggest one: no detail. No date, no amount, no invoice number, nothing to show the sender knows this particular client. That last tell does the most damage.

Don't get hung up on em dashes or any single word. The real lesson of the study is that the tells keep moving. Last year's giveaway gets trained out, and next year's model brings its own. Memorizing a list of banned words is a losing game. A better test is simple: read the message out loud and ask yourself whether you'd say it to this client face to face. If you'd never say “I wanted to kindly follow up regarding the outstanding balance,” don't send it. You'd say “Hi Dana, just checking you got the invoice from the 12th?” That version is shorter and specific, and nobody would mistake it for a template. That's exactly why clients answer it.

What do AI written messages look like before and after a fix?

AI written messages usually go wrong in the same way. They're longer, vaguer and more formal than the situation calls for, so fixing them mostly means cutting. Take a typical AI-drafted payment reminder: “Dear Dana, I hope you're doing well. I wanted to kindly follow up regarding invoice #142, which remains outstanding. Please let me know if there's anything I can do to assist with processing. Best regards.” Nothing in it is wrong, and that's the problem. It's polite, complete and easy to forget. Now compare: “Hi Dana, just checking you got the invoice from the 12th? Happy to resend if it went missing.” It asks for the same thing in about half the words, and it sounds like you.

The same pattern shows up in the messages small businesses send every week. A document request becomes “Hey Tom, could you send the signed lease by Thursday? The bank needs it.” A quote follow-up becomes “Hi Rachel, any thoughts on the kitchen quote? Happy to adjust the tiles if the price is off.” A rescheduling note becomes “Morning Avi, can we move Tuesday to 3pm? Something came up.” Look at what the human versions have in common: a first name, one concrete detail, one clear ask and often a reason. They don't apologize for getting in touch. They don't thank anyone in advance. And on WhatsApp, they never end with “Kind regards.”

Here's a less popular point: a message that's too casual is just as fake as one that's too stiff. Some people overcorrect by asking the AI to “sound human” and get chirpy messages full of emoji, “hey there!” and “no worries at all!!” If you've never sent a client a smiley face, a sudden burst of them will feel off. The goal isn't to sound human in general. It's to sound like you. If you normally write “Hi Mr. Levi” and sign with your name, keep doing that. Clients notice when you stop sounding like your usual self, because they've been reading your messages for months and know your voice better than you'd think.

How do you make AI messages sound human?

To make AI messages sound human, write the AI a brief that's as specific as you want the message to be. Say who it's for, the one thing you need, the detail that shows you know them, and how you normally talk. Most robotic drafts start with robotic prompts. “Write a follow-up to a client about an unpaid invoice” gets you exactly the bland template you'd expect. Try this instead: “Message Dana, who usually pays quickly, about invoice 142 from the 12th. Two lines, first name, no ‘hope you're well’, sign it Sagi.” That gets you something you'd actually send. Fifteen extra words in the brief save you from rewriting the whole draft afterwards.

A few rules of thumb make any brief better. Set a length limit, because “two sentences max” stops over-explaining faster than any other instruction. Ban the openers and sign-offs you never use, like “I hope this finds you well” and “Best regards.” Paste in one real message you've sent before so the AI can copy your rhythm instead of guessing. Describe the relationship, since a supplier you've worked with for years gets a different tone from a brand-new client. And say what you want back. “Ask her to confirm a payment date” is clearer than “follow up,” and it builds the whole message around a question the other person can actually answer.

Then edit hard. Delete the first sentence of almost any AI draft and it gets better, because that's where all the throat-clearing happens. Replace every vague noun with a concrete one: “the documents” becomes “the signed lease,” and “the matter” becomes “the leak under the sink.” Cut any word that's only there to sound nice, like “kindly,” “truly” or “just wanted to.” Read the message on your phone, not your laptop, because that's where your client will see it. If it fills the screen, it's too long. Last, keep a short list of the fixes you make again and again, and add them to your next brief so you stop correcting the same things.

Can an AI assistant follow up with clients without sounding like a bot?

Yes. An AI assistant can follow up with clients for you without sounding like a bot, as long as you see every message before it goes out and can set the tone in plain words. That's how outreach tasks work in Agent Sonic. They're in Beta on the Pro plan. You tell Sonic in WhatsApp who to message and what you need to find out, for example “Ask Dana when the invoice will be ready and let me know.” Before anything is sent, Sonic shows you the exact first message for each person. Reply “yes” to send it, “no” to drop it, or say what to change, like “shorter, no emoji” or “call me Sagi.”

Those tone notes aren't a one-off. They stay with the task and shape every message in that conversation, not just the first one, and the first message goes out exactly as you approved it. Sonic writes from its own WhatsApp number and always introduces itself as your assistant, so nobody thinks they're talking to you. That's the most honest answer to “can clients tell?” Sonic asks your question, follows up at most once if the answer is vague, and stops when it has the answer, the person declines or they ask it to stop. You get the result for each person plus a short summary, and you can check in anytime with “Did Dana answer yet?”

The limits are deliberate, and they keep things polite. You can run one outreach task at a time and message up to three people a day. First messages only go out between 08:00 and 21:00 in the recipient's time zone. Nobody gets contacted more than once a day. If someone doesn't reply, Sonic doesn't nag them. After 24 hours, it just tells you. Anyone who says “stop” gets a short apology and will never be messaged again by anyone using Sonic. Outreach works in one-to-one chats, not groups. The tutorial on how Sonic messages people for you walks through a full task, and you can get access to Sonic here.

Frequently asked questions