Short answer: use it for drafts, variations, and the blank-page problem. Don't let it decide what your business sounds like, and never let it write anything containing a claim. The line is not "AI bad, human good" — it's about which parts of the job depend on knowing things a model can't know.
The tell is not grammar
People assume AI writing is spotted by clunky phrasing. It usually isn't. Current tools write clean, confident sentences.
What gives it away is the absence of specifics. AI-written captions describe a category rather than a business. "Freshly prepared with quality ingredients." "Serving our community with pride." "Book today and see the difference." Every clause could belong to any of ten thousand businesses.
Yours knows things no model does. That the Tuesday lunch rush is nurses from the hospital two streets over. That you started doing repairs on Sundays because of one customer who couldn't get time off. That the bread order goes in at 5am and the reason the sourdough sells out by noon is that you deliberately under-bake the batch.
None of that is in the training data. All of it is why someone follows you.
Where it genuinely helps
I'd use it, and we do, for these:
- Beating the blank page. Ten caption angles for one photo, then you pick and rewrite. Far easier than starting from nothing.
- Reformatting for platforms. The same message as a punchy TikTok hook, a longer Instagram caption, and a Google Business post. This is mechanical work and it's fine.
- Second-draft tightening. You wrote something honest and rambling. Asking for it shorter, keeping the specifics, works well.
- Research scaffolding. Hashtag clusters, keyword variations, question lists worth answering. Verify before you use them.
Where it costs you
- Replies and DMs. This is where trust is either built or destroyed, and a generic response to a real complaint is worse than a slow one. Answer these yourself.
- Anything with a claim in it. Prices, ingredients, guarantees, turnaround times, "voted best in the county." A model will produce these fluently and some will be wrong. A confident false claim in your own caption is your problem, not the tool's.
- Your origin story, your values, your apology when something goes wrong. These are load-bearing. If they read as generated, everything around them gets doubted too.
The test worth using
Before a caption goes out, ask: could a competitor post this word for word tomorrow?
If yes, it says nothing. Not because a machine wrote it — plenty of human-written captions fail this too — but because it contains no information. Add one specific thing only you could know, and the same caption starts doing work.
That test also happens to be the best defence against sounding automated, without any need to detect anything.
The uncomfortable part
Your customers are being served more generated content every month, across every platform, from every business. The predictable result is that the generic register — confident, smooth, specific about nothing — is becoming an active negative signal. It reads as nobody was really here.
Which is quietly good news for a small business. The advantage was always specificity, and specificity is the one thing that doesn't scale. You have it and a content farm doesn't.
What we do
We draft with AI and finish by hand, and we don't publish anything about your business that you haven't told us. If a caption mentions your Tuesday special, it's because someone asked you what runs on Tuesdays.
That's slower. It's also the only version where the account still sounds like a person after six months — which is the point of doing any of this.
Ready to stop posting at midnight?
Book a 30-minute discovery call. We'll tell you straight whether a content batch makes sense for your business — and which tier fits.