How to tell if an AI caption tool sounds like you

Take the caption a tool wrote for you and picture it under a different stylist's photo of a different client. If it still fits, the tool wrote from a category rather than from your work. That is the whole test, and the other tells people list follow from it.

The reason this comes up at all is that the caption is the part that stalls. The photo took thirty seconds at the chair. Then the cursor sits in an empty box and the post does not go out.

What the research says, and where it stops

Bynder put a ChatGPT-written article and a copywriter-written one in front of 2,000 people in the US and the UK in March 2024, neither of them labelled. Among people with a preference, 56% picked the AI version as the more engaging one.^[2] The number worth building on came from the same study: 82% agreed with the statement "I don't mind if brands use AI to help write copy, as long as the piece feels like it is written by a human."^[1] The objection is not to the machine. It is to the result reading like one.

Sentiment has hardened since. Fractl surveyed 1,008 US consumers in the second quarter of 2026 and found 40% said their trust in a favourite brand would decrease if it used AI for most of its marketing, against 20% who said the same in 2025.^[3] The Nuremberg Institute for Market Decisions ran the label itself as an experiment, and found that identical ads presented as AI-generated were rated less natural and less useful than the same ads presented as human-made.^[4]

To be clear about what that covers: articles, ads, and national brand marketing. None of it looked at salon captions or at anything local, and somebody deciding whether to book with the barber four blocks away is doing something different from somebody reading a brand's blog. Treat the numbers as the shape of the reaction rather than its size in your feed.

The tells, named

It would fit any photo of that service. The swap test above. Paste the caption under three of your own older posts and count how many it suits.

An opening question nobody asked. "Ready for a change?" "Looking for your next look?" A caption that opens by addressing a stranger was written for a stranger.

Three adjectives where one would do. Gorgeous, radiant, stunning, stacked in one sentence. Stacking is what a writer reaches for with nothing specific to say about the head in the frame.

Trade vocabulary a client would not use. She says she wants to go lighter without looking blonde. She does not say dimensional lived-in bronde with a diffused root.

Emoji standing in for punctuation. Not the occasional one. The pattern where an emoji sits between every clause, doing the job a comma was going to do.

A fact nobody checked. A price, a turnaround time, an opening on Thursday. This is the only tell that costs money: a caption inventing availability produces a phone call somebody has to walk back.

The same rhythm every time. Setup, transformation word, call to action. Unremarkable once. Recognisable by the fourth post, because every caption started from the same blank prompt.

What a tool has to be reading

Two inputs, and both are answerable before you pay for anything.

The first is the photo itself. A caption written from what is in the frame can say something that is only true of that frame, which is what stops a balayage shot getting a caption about fresh cuts. A tool that never looks at the image is working from a service name and the word salon, and no amount of tone setting repairs that.

The second is your own past posts, and this is the one people skip. Ask whether the tool reads what your account has already published, or whether it asks you to describe your voice on a form. A form gives it a description of your voice. Your archive gives it a sample. Those produce different sentences, and the gap opens up around the fourth caption, when a described voice starts repeating its own summary of itself.

The two sentences a tool cannot write

I built SalonFawn, so grain of salt on this part. Captions are written from what is actually in the photo, and voice learning studies the account's own best-performing posts and the edits you make, then writes in that voice rather than a generic one. That is the mechanism, and it is the same one I would hold any other tool to.

Here is what it will not do. It will not write about anything that is not in the photo. It does not know she had been growing that out for two years, or that this was her first appointment since she moved here. The AI acts as an editor and a scheduler, never a generator. It never invents facts about services, prices, or availability, and it never generates footage.

The repair, when a caption still does not sound right, is two sentences only you have the information to write. What she asked for, in her words. What you did about it. "She wanted to go lighter without going blonde, so we lifted the mid-lengths and left the root alone" is a caption. Everything a tool produces is scaffolding around that.

If you find yourself editing the same thing out every week, the repetition is itself the signal. A tool that does not learn from your edits will hand you the same sentence next Tuesday.

Shooting the photo well sits upstream of all of it: a caption written from the frame is only as specific as the frame. How to shoot a before and after that reads at thumbnail size covers that half. The capture guide lists what is worth grabbing while the phone is already out. Your call on the tool.