Read ten Google review replies from ten different small businesses and you'll notice something odd: they could all have been written by the same person. "Thank you so much for your kind words! We're thrilled you had a great experience and can't wait to welcome you back!" Different shop, same sentence, over and over.
If you've been using AI to draft these (replies, social captions, the odd newsletter line), you already know why. You typed three words, it filled in the rest, and the rest sounds exactly like everyone else's rest. That's not something you're doing wrong. It's how the tool works. And lately, it's started costing more than it used to.
The sameness isn't an accident
There's a good technical reason AI text drifts toward the same handful of phrases, and it's worth understanding because it changes what you do about it. Sophie Alpert, a researcher who's written carefully about this, put it plainly: every rewrite or rephrase changes the meaning of what you were trying to say, and if the entity doing the rewriting doesn't have your actual mental picture of what happened, something gets lost.
Think about what that means for a review reply. You know which review you're answering. You remember the customer, maybe what they said, maybe what actually happened in the shop that day. The AI knows none of that unless you tell it. So when you ask it to "write a thankful reply," it can't reach for your specific memory. It reaches for the statistically safest words instead: thrilled, delighted, can't wait to see you again. Those words aren't wrong. They're just not yours, and they're not about this particular customer. Do that forty times and you get forty identical replies.
Customers and platforms have both started noticing
This used to be a minor cosmetic issue. It isn't anymore. A 2025 Gartner survey found 53% of consumers now distrust AI-generated search results and summaries outright, and a separate global survey by Baringo put the number who feel uncomfortable with AI-generated content at 70%. Businesses relying on generic AI text for search-facing content have seen organic traffic drop by 5 to 35%, because both readers and search engines increasingly treat generic phrasing as a signal to skip past, not read.
Platforms are adjusting too: LinkedIn has said it will actively deprioritize "generic" AI-flavoured posts even as it rolls out its own AI writing tools, and in the EU, new rules due at the end of this year will start requiring some AI-generated content to carry a visible disclosure. None of this means AI-written text is banned or bad. It means the free ride, where "AI wrote it and nobody could tell the difference," is ending. The difference is starting to show, and people are starting to look for it.
A drafting tool, not a mind reader
It helps to be precise about what's actually happening when you ask an AI to write your review reply. It isn't reading your mind or acting on your behalf the way a true AI agent would (something we explained in What is an AI "agent", actually?). It's a one-shot text generator: you give it a prompt, it gives you back its best statistical guess at what that prompt usually leads to. No memory of your shop, no access to what actually happened, nothing to draw on but the words you typed. The generic output isn't the tool failing: it's the tool doing exactly what a prompt with no specific detail in it will always produce.
The one-line fix
Here's the part that actually saves the reply: before you send anything the AI drafted, add one detail it couldn't have known. Not a rewrite, one sentence. If a customer mentioned your dog in their review, mention the dog back. If they waited for a part you had to order in, say so: "glad the part came in before your trip." If nothing distinctive happened, say what you're actually thinking: "means a lot, thank you" is shorter and truer than "we're thrilled you had a great experience."
That one sentence is doing something specific: it's the piece of information only you had, so it's the piece the AI could never have generated on its own. It's also, not coincidentally, the part a customer actually reads.
Honest limits
Don't turn this into a new task. If you're replying to forty reviews on a Sunday evening, a warm-but-generic reply to review thirty-nine is still better than no reply at all. The fix is worth applying where you have a genuine detail handy, not something to chase for every single line. And don't go hunting for "AI detectors" or obsess over sounding perfectly human; reliable AI detection doesn't really exist, and chasing it is a worse use of your evening than the replies themselves. Specificity is the actual goal. Sounding human is just what specificity happens to look like from the outside.
One thing to try this week
Next time you use AI to draft a reply, caption, or short message, before you send it, add exactly one detail the AI didn't have. Do it for your next five. You'll likely notice it takes seconds, and you'll definitely notice the reply doesn't read like the other nine businesses' anymore.
Clara F.