An AI caption almost never reads as broken. That is exactly the problem. It reads as fine, and fine is what makes people scroll past.
Language models are trained to produce the most probable next word. The most probable phrasing is, by definition, the phrasing everyone else is already using. So the output is smooth, grammatical, and interchangeable with a thousand other brands writing about the same thing.
The good news: machine-written text leaves fingerprints, and once you know them you cannot unsee them. Here are the twelve that matter most for social captions, why each one costs you attention, and how to fix it.
1. Stock phrases you would never say out loud
The clearest signal. Certain phrases appear far more often in AI text than in writing by people: "dive into", "elevate your", "unlock the power of", "game-changer", "in today's fast-paced world", "the possibilities are endless", "stay tuned".
None of them are wrong. They are just worn smooth from overuse, and they carry no information.
Fix: Read the sentence aloud. If you would not say it to a customer standing in front of you, cut it. "Unlock the power of our new tool" becomes "Here is what the new tool does".
2. Clustered em dashes
People use commas and full stops. Language models reach for the em dash, and they reach for it repeatedly. Two or three in a short caption is one of the most reliable tells there is.
Fix: Replace most of them with a comma or start a new sentence. If a caption has three em dashes, you almost certainly want three shorter sentences instead.
3. The "not just X, but Y" construction
"It's not just a tool, it's a journey." "More than just software." This is a language-model staple. It sounds like insight while committing to nothing.
Fix: Delete the first half and keep the claim. "It's not just a tool, it's a workflow" becomes "It replaces four steps in your workflow", which is a statement someone can actually check.
4. Rule-of-three stacking
Three adjectives, three benefits, three clauses, over and over. "Fast, simple, and reliable." "Plan, create, and publish." One triple is rhetoric. Three triples in a row is a rhythm no human writes by accident.
Fix: Keep one, cut the rest. Uneven sentence lengths are what make writing sound like a person.
5. No numbers, names, or moments
The surest sign of interchangeable text is that nothing in it could only be true of you. No figure, no client name, no specific Tuesday when the thing happened.
Fix: Add exactly one concrete detail. "Our clients see great results" becomes "One client cut their approval time from four days to one". Specificity is the cheapest credibility there is.
6. An opener that earns nothing
"In today's digital world." "We are excited to announce." "It's that time again." These openings ask for attention before giving a reason to grant it, and the first line is the only line most people read.
Fix: Lead with the most surprising or concrete thing you have. If the interesting part is in sentence four, move it to sentence one.
7. Sentences that run past 28 words
AI defaults to long, evenly weighted, subordinate-clause-heavy sentences. On a phone, in a feed, those are where people stop reading.
Fix: Break anything over roughly 25 words into two. Then check that not all the results are the same length.
8. Saturated hashtags
#love, #instagood, #photooftheday, #viral, #fyp. These tags are so crowded that a post appears and disappears within seconds, so they deliver essentially no reach. They also read as automated, because they are what a generator picks when it has nothing specific to say.
Fix: Trade reach for relevance. Five tags a niche audience actually follows will beat fifty generic ones every time.
9. The wrong number of hashtags for the platform
Hashtag counts are not universal. Facebook rewards almost none. LinkedIn and Instagram sit comfortably around three to five. A caption carrying thirty tags was not written for the platform it is on.
Fix: Match the platform. The table further down has the numbers.
10. Emoji volume that ignores the channel
A caption with an emoji after every clause reads as mass-produced. And the right amount differs by platform: what looks lively on Instagram looks unserious on LinkedIn.
Fix: Use emoji where they replace a word or set a tone, not as decoration between every phrase.
11. Length that ignores where the platform cuts
Every feed truncates. Instagram shows roughly the first 125 characters before "more", Facebook around 250, LinkedIn about 200, TikTok around 100. A caption whose point lands in character 300 is a caption most people never reach.
Fix: Put the hook and the payoff before the cut. Treat everything after it as the people who already decided to stay.
12. No reason to do anything
AI drafts describe. They rarely ask. Without a next step, a caption that was read still produces nothing.
Fix: End with one specific action. Not "let us know your thoughts", which nobody does, but "tell me which of the two you would pick".
What each platform actually expects
These are the values worth checking a caption against, and they are the same ones our own checker scores against.
| Platform | Sweet spot | Visible before "more" | Hashtags | Emoji |
|---|---|---|---|---|
| 125 to 400 characters | 125 | 3 to 5 | up to 5 | |
| 40 to 250 characters | 250 | 0 to 2 | up to 3 | |
| 200 to 1300 characters | 200 | 3 to 5 | 0 to 2 | |
| TikTok | 30 to 150 characters | 100 | 3 to 5 | up to 4 |
A worked example
Here is a caption with most of the patterns above:
In today's fast-paced digital world, we are thrilled to unlock the power of seamless content creation. It's not just a tool, it's a journey. Our cutting-edge, intuitive, and robust platform revolutionizes the way brands communicate. Dive in and elevate your content! #love #instagood #viral #marketing #business #content
Stock phrases, a "not just" construction, rule-of-three stacking, no concrete detail, a weak opener, saturated hashtags. Now the same message with the patterns removed:
We rebuilt our editor last month. Approval time for one agency dropped from four days to one. Same team, same volume, one less bottleneck. Which part of your approval process eats the most time? #ContentOps #AgencyWorkflow #SocialMediaManagement
Shorter, more specific, and it says something only this company could say.
Does this mean you should stop using AI?
No, and that would be the wrong lesson. AI is genuinely good at the blank page, at variations, and at adapting one message to four platforms. What it is not good at is knowing which detail from your week is the one worth telling.
The workable split: let AI handle structure and speed, and supply the specifics yourself. The patterns above are what appear when nobody supplied any.
Summary
Three things worth remembering:
- AI captions fail by being average, not by being wrong. You are not looking for errors, you are looking for the absence of anything specific.
- One concrete detail fixes more than any rewrite. A number, a name, or a moment does more for credibility than swapping adjectives.
- Platform fit is half the battle. The same text is a good Instagram caption and a bad TikTok one, and length is usually the reason.
Want a second opinion on a specific caption? Our free AI caption checker scores any text against all twelve patterns above, shows you which lines trigger them, and returns a version rewritten for the platform you post on. No signup for the score and the rewrite.
Frequently Asked Questions
Can AI detectors reliably tell whether a caption was written by AI? Not with certainty, and you should be sceptical of any tool that claims otherwise. General-purpose detectors produce both false positives and false negatives, and short social captions give them very little to work with. What is reliable is looking for the specific patterns above: they tell you whether a caption reads as generic, which is the thing that actually costs you engagement.
Does it matter if my captions look AI-written? It matters when it makes you interchangeable. Nobody penalises you for using AI, but a caption built from the most probable phrasing sounds like a thousand competitors, and readers scroll past without consciously knowing why.
Do the social platforms punish AI-generated captions? There is no evidence that Instagram, Facebook, LinkedIn or TikTok demote a post for being AI-assisted. The damage is indirect: generic text earns fewer saves, comments and shares, and those are the signals that do affect reach.
How many of these patterns are too many? One or two in a caption is normal, and even good writing contains them occasionally. The problem is accumulation. When four or five show up in the same short text, the caption stops sounding like anyone in particular.
What is the fastest single fix? Add one concrete detail and cut the opening sentence. Most AI drafts warm up for a line and a half before saying anything, and most lack a single fact that could only come from you. Fixing those two things changes the whole caption.