Table of Content
I have written captions with AI on almost every working day for the past two years, across a dozen small brand accounts. The first draft is always fast. It is also, almost always, obvious. Not to me at first, but to the people scrolling. Somewhere around post three hundred I started keeping a private list of the phrases my tools reached for again and again, and that list is the reason this guide exists.
Here is the uncomfortable part. The problem is rarely the tool. Hand the same prompt to ChatGPT, Claude, or Gemini and you get competent, plausible, on-topic copy that sounds like every other brand feeding in the same prompt. The distance between “plausible caption” and “sounds like us” is the part you close by hand, and most people never learn to close it quickly.
What follows is the shortcut I wish someone had handed me on day one. It covers what gives AI captions away in 2026, the fixes that actually hold up, and a repeatable setup that gets your first draft close instead of generic. No theory for the sake of it. Everything here I have used on real accounts that needed to sound human by Monday morning.
Readers got fast at spotting machine copy
Audiences learned the patterns, and their patience ran out. A few numbers worth keeping in view before you post anything:
52% of consumers stop reading the moment they suspect a caption was written by AI. BYNDER | 13% say they fully trust AI-generated brand content. KLAVIYO | 87% already assume brands use at least some AI in their posts. CASHEW RESEARCH, 2026 |
Search engines watch the same patterns. Google's Information Gain signals reward pages and posts that avoid predictable AI phrasing and add something only you could write. So sounding human is a trust play and a ranking one at the same time.
The tells that give a caption away
These are the patterns editors and audiences flag in 2026. Read the middle column out loud and you will hear why they land wrong. The right column is what I swap in.
| THE TELL | HOW IT READS | THE HUMAN FIX |
|---|---|---|
| Em dashes as a habit | Great coffee — made slowly — for people who care. | Two short sentences, or one comma. |
| The contrast reframe | “It's not about the beans. It's about the ritual.” | Say the real point once, plainly. |
| Rule-of-three stacking | “Bold, bright, unforgettable.” | Keep one strong word. Drop the other two. |
| Questions no one answers | “Ever wonder what sets great coffee apart?” | Open with a specific detail instead. |
| Buzzword crutches | elevate, unlock, seamless, game-changer, delve | Name the concrete thing you mean. |
| Fake-profound filler | “quietly”, “the pull of”, “hold space”, “hits different” | Describe the actual feeling or action. |
| Flat emotional temperature | Correct, polished, and completely neutral. | Add a real reaction: annoyance, delight, a stance. |
| Announcement phrases | “Here's the best part.” “But here's the thing.” | Delete the drumroll. Let the line stand alone. |
Fix the voice before you edit
Most robotic output comes from a lazy loop: describe photo, get caption, post. The repair is front-loading context so the first draft starts close to how you actually talk. Here is the routine I run on every account.
1. Write a voice brief once
A short paragraph you paste at the top of every prompt. Treat it like a character sheet for how your brand talks: the tone, who the audience is and what they already know, sentence length, whether humor is welcome, and two or three things you would never say. You write it once and reuse it forever.
2. Feed it five to ten real examples
Paste five to ten of your best existing captions and tell the model to copy the rhythm, vocabulary, and opening patterns, not the topic. That single instruction, match the voice not the subject, does more work than any adjective like friendly ever will. Mix the formats a little too: a couple of posts, a line from an email, a product blurb.
3. Ask for ten, keep one
Never request a single caption. Ask for ten short variations, choose the one line that sounds most like you, and fold the best bits of the others into it. AI is a strong variation engine and a weak final author. Use it for the first, not the second.
4. Add one detail only you know
The regular who orders the same thing every Tuesday. The exact hour the light hits the counter. A model cannot invent the true specifics of your world, and those specifics are what make a caption unmistakably yours. One real detail beats a paragraph of adjectives.
5. Read it out loud
If you stumble on a word, or you would never say the line to a friend across a table, cut that part. Reading aloud catches the textbook-perfect phrasing that looks clean on screen and sounds wrong in the mouth. It is the fastest edit in this whole guide.
COPY-PASTE: THE VOICE BRIEF TEMPLATE Write in the voice of [who your brand sounds like]. Example: a blunt, funny barista who takes coffee seriously and themselves lightly. Audience: [who they are + what they already know] Rules: 1 to 3 sentences. Contractions on. Dry humor welcome. Never say: [your banned words, e.g. elevate, unlock, journey] Match the rhythm of the examples below, not the topic: [paste 5 to 10 of your real captions] |
The swap file
Keep a running list like this and paste it into your never-say line. When a model reaches for the left column, you already know the human version.
| WHAT THE MODEL REACHES FOR | WHAT A HUMAN WOULD WRITE |
|---|---|
| elevate your mornings | make mornings less of a slog |
| unlock your potential | get the thing you came for |
| a seamless experience | nothing to fuss with |
| this is a game-changer | this actually changed how we work |
| let's delve into | let's look at |
| a curated selection | the stuff we actually like |
| embark on a journey | start |
| in today's fast-paced world | (delete the whole phrase) |
One caption, fixed
Same photo of a morning brew. The first draft is what the model handed me. The second is what went out.
BEFORE • RAW AI DRAFT “Elevate your morning ritual with our seamless, hand-crafted brew, because it isn't just coffee, it's an experience. Ready to unlock your best day?” | AFTER • EDITED FOR VOICE We roast this one dark on purpose. It's the cup we make before the shop opens, when it's quiet and nobody's talking yet. Yours is ready at 7. |
What changed: no buzzwords, no contrast reframe, no rhetorical question. In their place: one true detail (before the shop opens), a real point of view, and a plain instruction at the end. It went from a template to something a person would actually say.
Keep this, cut that
If you remember nothing else, remember this split. The left side is what earns trust. The right side is what quietly erases it.
Keep ✓ Contractions, always ✓ One concrete detail per caption ✓ Sentence length that varies ✓ A clear opinion or reaction ✓ Plain verbs over fancy ones ✓ Your own recurring jokes | Cut ✕ Em dashes on autopilot ✕ Stacked adjectives ✕ Questions no one would answer ✕ “It's not X, it's Y” ✕ Buzzword verbs ✕ The same three words in every post |
The pre-publish check
Run this on every caption. It takes under a minute once it becomes a habit.
☐ Read it aloud once. Any stumble gets cut.
☐ Scan your last ten captions. Any word repeating three or more times? Swap it.
☐ One detail only you could know is in there. If not, add one.
☐ Zero rhetorical questions a real reader would never answer.
☐ No em dash you couldn't replace with a period or a comma.
☐ You would text this to a friend without cringing.
THE VERDICT
Two years in, and more edited drafts than I care to count, my honest take is simple. AI is the best first-draft partner I have ever worked with and a genuinely bad final author. Used as a ghostwriter, it produces the flat, plausible copy that readers now scroll straight past. Used as a fast sparring partner, with a voice brief, real examples, and one detail only you know, it saves me hours a week without costing me my voice.
The tools keep changing, and so do the tells. My ban list this month looks nothing like last year's, and yours will move too. What does not move is the habit: front-load your voice, ask for many, keep one, add a human detail, and read it aloud. Do that consistently and nobody scrolling will care whether a machine helped write it. That was always the point.
So start your swap file today, and add to it the moment a phrase starts turning up everywhere. The people who still sound human a year from now are the ones paying attention right now. Be one of them.