Table of Content
I write captions for a living, and I started leaning on AI drafting tools the week they got good enough to be useful. The first fortnight was a small revolution. A blank content calendar filled itself in an afternoon. By the second month, something felt off. I was scrolling my own feed and could not tell which posts I had written and which the machine had. Then a client said the thing no copywriter wants to hear: “These sound like everyone else's now.”
She was right. The tools had not made my writing worse, exactly. They had made it average, which for a brand is worse than bad. So I spent the next year figuring out precisely where AI captions go wrong and how to pull them back. This is that playbook. No theory you have read in ten other posts, just the patterns I can now spot from three posts away and the fixes that actually land.
Why AI captions sound like AI
You cannot fix the symptom until you understand the machine underneath. A language model writes by predicting the most probable next word, over and over. That single fact explains almost everything you dislike about its output.
REASON A It reaches for the average Probable means common. The model gravitates to the phrasing millions of other pages already used, which is the opposite of a distinct voice. | REASON B It was trained to be safe Helpful and inoffensive was the goal. Safe writing hedges, balances both sides, and never risks a real opinion. Captions live and die on opinion. |
REASON C It over-explains The model was rewarded for being thorough, so it sets up, contextualizes, and concludes. A good caption skips the setup and starts at the interesting part. | REASON D It has no Tuesday It never had a bad commute, a favorite mug, or a grudge against slow WiFi. Specific lived detail is where human writing gets its texture, and the model has none. |
| The machine is not trying to sound generic. Generic is simply what “most likely” looks like when you average a few billion sentences. Your job as the editor is to drag it away from the average on purpose. |
The words and patterns that give it away
After enough drafts you build a mental blocklist. These are the exact phrases and structures that make a reader's guard go up, even when they cannot name why. Treat this table as a search-and-destroy list before anything goes live.
| THE TELL | WHY IT READS AS MACHINE |
|---|---|
| elevate / unlock / unleash | Ambition verbs with no object a human would actually use. Nobody unlocks a latte. |
| game-changer | A claim with zero evidence attached. Real people name the specific thing that changed. |
| seamless / effortless | Marketing filler. It describes a feeling the reader is supposed to have, not a thing that happened. |
| dive in / let's dive in | A transition to nowhere. It announces content instead of being the content. |
| in today's fast-paced world | The single most exhausted opener online. An instant signal that no human sweated over this line. |
| the realm of / a tapestry of | Reach-for-grandeur nouns. They inflate ordinary subjects and fool no one. |
| a testament to | Faux-gravitas that models deploy on anything from skincare to sourdough. |
| not only / but also | A tidy symmetrical frame that real speech almost never falls into naturally. |
| when it comes to | Four words of runway before the actual point. Cut them and start at the point. |
| emoji bookends | Emoji used as symmetrical decoration rather than as a reaction a person would have. |
Beyond the vocabulary
Word swaps are the easy half. The deeper tells are structural, and readers feel these even more strongly than the buzzwords.
PATTERN 01 The rule of three, everywhere “Fresh, bold, unforgettable.” Models love triplets. One or two well-chosen words almost always hit harder. | PATTERN 02 Suspicious symmetry Every sentence the same medium length, every clause balanced. Human rhythm is lumpy. It stutters, then runs long. |
PATTERN 03 The rhetorical question opener “Ready to transform your mornings?” Nobody asks this out loud. It is a template disguised as a hook. | PATTERN 04 Hashtags that restate the caption Tagging a coffee post with coffee, coffeelover, morningroutine adds nothing. It is filler pretending to be reach. |
Eight moves that make a caption sound human
Deleting tells gets you to neutral. These techniques get you to a voice. I reach for at least two of them on every caption, and the best posts usually use four or five without me planning it.
| 1 | Lead with one concrete detail Name the place, the price, the exact time. “The $4 latte at the corner spot on 6th” beats “a delicious coffee” because specificity is proof a real person was there. |
| 2 | Hold one opinion you would defend Say something a reasonable person could disagree with. Neutral is invisible. A mild, honest take gives the reader something to nod or argue with. |
| 3 | Vary your sentence length on purpose Follow a long, winding sentence with a short one. Then a fragment. That uneven rhythm is one of the strongest human signals you can send. |
| 4 | Cut the first sentence The model's opener is almost always throat-clearing. Delete it and start on line two, where the actual story tends to begin. |
| 5 | Write it the way you would text a friend Use the small words: kind of, honestly, so. Contractions everywhere. If you would never say a phrase out loud, do not post it. |
| 6 | Leave one deliberate imperfection A sentence fragment. A slightly odd word choice. A thought that trails into a joke. Machines polish everything smooth, so a small rough edge reads as a fingerprint. |
| 7 | Reference something only your people would get An inside joke, a local landmark, a running bit from last week's post. Shared context is the fastest shortcut to sounding like a real account. |
| 8 | Anchor to a moment, not a concept “Wellness” is a concept. “I skipped my run twice this week and felt it on Friday” is a moment. Moments carry the emotion that concepts only gesture at. |
Real rewrites, marked up
This is the part I wish someone had shown me a year ago. Same brief, same product, one draft left as the tool produced it and one after five minutes of editing. The struck-through text is what I cut and why it had to go.
AI DRAFT Elevate your morning routine with our seamless, game-changing blend that unlocks your full potential. ☕✨ #coffee #coffeelover #morningmotivation |
HUMAN REWRITE I have had the same $4 oat latte every Tuesday for a year. My barista knows my order and my dog's name. That is the entire pitch. |
AI DRAFT In today's fast-paced world, finding time for self-care can be a challenge. That's why we've curated the ultimate skincare routine to help you glow from within. 🌿✨ |
HUMAN REWRITE I did a three-step routine for a month straight, which is a personal record. My skin is fine. My real win is that I finally stopped buying products I never open. |
AI DRAFT Ready to take your workouts to the next level? Our new program is a testament to hard work and dedication, designed to help you unleash your inner athlete. 💪🔥 |
HUMAN REWRITE Week one of the new program: I could barely finish. Week six: I finished, then went for a walk after. Nobody warns you that the walk is the actual flex. |
| Notice what the rewrites share. A number, a timeframe, one small confession, and zero ambition verbs. None of them are clever. They just sound like a person who was actually there. |
My six-step de-robot edit
I still use AI to beat the blank page, because staring at nothing is its own kind of tax. The value is in what happens after the draft. Here is the pass I run on every single one, and it takes under five minutes once it is a habit.
| 1 | Delete the opening line It is nearly always setup. Read what is left. The post usually starts stronger without it. |
| 2 | Run the blocklist Scan for elevate, unlock, seamless, dive in, and every tell from the table above. Kill them on sight. |
| 3 | Trade one abstraction for a thing Swap an abstract noun like quality or journey for something you can see, touch, or price. Concrete beats grand every time. |
| 4 | Break the symmetry Find two sentences of similar length and shorten one hard. Two words is fine. The jolt is the point. |
| 5 | Add a real number, name, or time A price, a day of the week, a first name. One verifiable detail does more for trust than any adjective. |
| 6 | Read it aloud, cut what you would never say This last step catches everything the other five missed. If a line feels awkward in your mouth, it reads awkward on the screen. |
The pre-post checklist
Screenshot this or pin it near your workspace. If a caption cannot check most of these boxes, it goes back for another pass.
Before you hit publish ✓ There is one concrete detail: a name, price, place, or time. ✓ There is one opinion a reader could disagree with. ✓ The sentences are not all the same length. ✓ Not a single word from the blocklist survived. ✓ It contains at least one phrase I would actually say out loud. ✓ The first sentence earns its place, or it is gone. ✓ The emoji are reactions, not symmetrical decoration. ✓ The hashtags add reach, not a restatement of the caption. |
A note on where AI still earns its keep
None of this means the tools are useless. They are excellent at the parts of the job that are not the writing. Give me twenty hook variations to react to, a rough outline for a carousel, a first pass at trimming a long caption to fit a character limit, or ten alternate calls to action, and I will take all of it. The mistake is treating the draft as the finished thing. The draft is raw material. You are still the one who has to make it sound like someone.
Things people ask me about this
These four come up in almost every workshop I run, so they are worth answering directly.
Q Do AI detector tools actually catch captions? Not reliably, and I would not build a process around them. They flag long, uniform text, and captions are short and choppy. Trust your own ear over a score. | Q Will editing this heavily hurt my reach? The opposite, in my experience. Posts that sound like a real person tend to earn more saves and replies, and those signals matter more to most feeds than keyword stuffing. |
Q What if I do not have a personal detail to add? You almost always do once you look. A price, the day you shot the photo, how long it took, a small thing that went wrong. Reporting beats inventing, so go find the real detail. | Q Can I train the tool to sound like me? Partly. Feeding it ten of your best past captions as examples helps it copy your rhythm. It still cannot invent your lived context, so the final edit stays on you. |
THE VERDICTUse it for the blank page. Own the last mile yourself. Fourteen months in, my honest take is that AI is a genuinely good drafting partner and a genuinely bad final-draft writer, and pretending otherwise is how brands end up sounding like each other. The tools got me past the hardest part of the day, which is starting. What they could never do was remember that my client's audience jokes about their office cat, or that a $4 detail lands harder than a paragraph of adjectives. The tells in this piece are learnable in a week. Once you can see them, you cannot unsee them, and that is the whole game. Let the machine hand you the clay. Then sit with it for five minutes and put your thumbprints all over it. That five minutes is the difference between a caption that scrolls past and one that sounds like it came from a person worth following.
|