You've felt it. You're scrolling, and a caption stops you — not because it's good, but because it's too tidy. Three neat clauses. A rhetorical question. A closing line that sounds like a motivational poster wrote it. Something's off, and you can't quite name it.
I've spent a lot of time reading AI-written captions side by side with human ones, and the honest summary is this: the tells are real, but most of the ones people repeat online are wrong. Emoji count tells you nothing. Perfect grammar tells you nothing. What actually gives it away is rhythm.
The short version
AI captions are unusually evenly paced. Human captions are lumpy — a long rambling sentence, then three words, then a tangent nobody asked for. If every sentence is roughly the same length and every paragraph lands the same way, that's your signal.
The tells that actually work
These are patterns I keep seeing across Instagram, LinkedIn and X. None of them is proof on its own. Two or three together, though, and you're usually looking at a language model.
1. The “It's not X. It's Y.” construction
This one is everywhere. “It's not about the destination. It's about the journey.” “This isn't a setback. It's a setup.” Models love this shape because it sounds profound and costs nothing. Humans use it too — but roughly once a year, not twice per caption.
2. Everything comes in threes
Three adjectives. Three bullet points. Three-part closing line. The rule of three is genuinely good writing advice, which is exactly why models over-apply it. When you notice a caption where every list happens to have three items, that regularity is doing more work than the writer intended.
3. Suspiciously balanced sentences
Read the caption out loud. Human writing stumbles — a clause runs long, a thought gets abandoned halfway, someone forgets to close a parenthesis. AI writing scans almost like verse: each sentence about fifteen to twenty words, each one grammatically complete, each one ending on a beat. That evenness is the single most reliable signal we measure.
4. Hashtags that don't match the post
A person writing about their sourdough starter tags #sourdough #baking #kitchenfail. A model generating a post about sourdough tags #sourdough #baking #foodie #inspiration #community #journey — broad, generic, and reaching for engagement rather than describing anything. Mismatched abstraction between the caption and the tags is a decent tell on Instagram and TikTok specifically.
5. The LinkedIn opener
LinkedIn deserves its own entry because the genre has collapsed into a formula: a one-line hook, a line break, a short dramatic sentence, another line break, then a story that resolves into a business lesson. “I got rejected 47 times.” Line break. “Here's what it taught me.” Models learned this structure from a million real posts, so they reproduce it perfectly — which, ironically, is what makes it detectable. Real people break their own formula. Models don't.
The tells that don't work (please stop repeating these)
This section matters more than the last one, because acting on a bad tell means accusing real people of things they didn't do.
“It uses em dashes, so it's AI”
No. The em dash became a meme tell in 2025 and it was never reliable. Plenty of good writers use them heavily — and current models have been tuned away from them precisely because people started noticing. If anything, heavy em-dash use now correlates slightly with human writing.
“The grammar is perfect”
So is the grammar of anyone who typed their caption in a notes app and read it twice. Perfect grammar is a proxy for care, not for authorship. This tell punishes careful writers, non-native speakers who over-edit, and anyone whose phone autocorrects aggressively.
“It uses fancy words like 'delve' or 'tapestry'”
There's a kernel of truth here — certain models did overuse a handful of words — but it decays fast. Word-level tells have a shelf life of about one model release. Structure lasts much longer, which is why detection tools weigh rhythm and predictability over vocabulary.
It looks different on each platform
The same model produces very different-looking output depending on where it's posting, because the prompt usually includes the platform.
Captions trend long and story-shaped. The tell is usually a mismatch: a highly polished, structured paragraph attached to a spontaneous-looking photo. People who write casually don't suddenly produce a five-sentence narrative arc about a coffee.
X (Twitter)
Short posts are genuinely hard to judge — there simply isn't enough text. Where AI shows up clearly is in threads. Watch for numbered posts that each hit the same length and each end with a hook into the next one. Humans lose steam around post four. Models don't.
TikTok
The caption is often an afterthought, so it's thin either way. The more useful signal is the on-screen text or the script, if you can get it — scripts are where the generated structure shows up.
The highest density of AI-written content of any major platform, and also the hardest to call, because the human posts imitate the same formula. Weight the specifics: real posts name real companies, real numbers, real people. Generated ones stay abstract.
Why anyone actually needs to check this
Not curiosity, mostly. The people who ask us tend to have a concrete reason:
- Hiring. A recruiter reading a candidate's LinkedIn wants to know whose voice they're reading.
- Brand partnerships. Agencies vetting a creator before paying them to write want to know if the “voice” they're buying is a prompt.
- Journalism and research. Anyone quoting a post needs to know whether a real person said it.
- Moderation. Coordinated engagement farming is generated at volume, and it reads like it.
Notice what's missing from that list: catching a friend. If you're about to confront someone over a caption, the honest advice is don't. The tools — ours included — give you a probability, not a confession.
Checking a post without reading tea leaves
If you'd rather not eyeball it, you can paste a public post link into our social media AI detector and it will pull the caption and score it, showing the specific sentences that pushed the number up. It works with Instagram, X, TikTok, Threads, Reddit, LinkedIn, Facebook and YouTube.
Two things worth knowing before you use it, or any tool like it. First, it analyzes the text. It shows the post's image so you know it read the right post, but it does not score the image — picture forensics is a different problem and we'd rather say so than hand you a number we made up. Second, it refuses to score captions under about twenty words. Below that there isn't enough signal, and any percentage would be closer to a coin flip than a measurement.
The part most articles skip
AI detection is probabilistic. Always. A high score means the writing carries patterns strongly associated with language models — it does not mean someone lied to you. Some people naturally write in clean, even, well-structured prose. Marketing copy has sounded faintly machine-generated since long before machines could generate it. And anyone can run a human-written caption through a rewriting tool and flip the score.
So use the score the way you'd use a smoke alarm: as a reason to go look, not as a verdict. The flagged sentences matter more than the percentage. If the reasoning doesn't convince you, trust your own read — you know the person's voice better than any model does.
The short answer
Stop counting emoji and em dashes. Read for rhythm instead. Human writing is uneven, specific and occasionally a mess; generated writing is smooth, balanced and strangely reluctant to name anything concrete. Once you've noticed that, you'll see it everywhere — which is either a useful skill or a curse, depending on how much time you spend online.
