AI generated content sounds like AI for one reason: the sentences have no facts in them, so the model fills the space with shape instead. Eight habits produce almost all of it. Throat-clearing openers, tricolons, "not only... but also", hedging, stacked adjectives, em dash overuse, generic sensory claims, and calls to action that ask for nothing.
None of them get fixed by telling the model to sound human. "Human" is an adjective, and a model cannot check itself against an adjective. Every fix below is a rule instead: something with a number in it, a banned string, or a required input. Rules survive the next prompt. Vibes do not.
Why AI copy is detectable at all
Two things cause it. The first is the training objective. A language model is very good at producing the most likely next sentence, and the most likely sentence about a coffee brand is the average of every sentence ever written about a coffee brand. Average writing is not bad writing. It is writing that could be about anyone, which on a social feed amounts to the same thing.
The second cause is the brief. Most prompts ask for a caption about a product and supply no facts about that product. The model returns a full caption anyway, because it always returns a full caption. With nothing specific to say it reaches for rhythm, hedges and adjectives, which are the linguistic equivalent of a placeholder image.
The tells are not style problems. They are what a sentence looks like when there is nothing inside it.
Family one: tells that fill empty space
Throat-clearing openers
"In today's fast-paced digital landscape." "When it comes to skincare." "Let's be honest." The first sentence is a warm-up, and warm-ups are high-probability text because they commit to nothing. They are always available, whatever the subject.
The structural fix is a rule about the first sentence, not a request for a better one. Require the opening sentence to contain at least one of these: a number, a place, a date, a proper noun, or a price. Then run the delete test. Remove the first sentence. If the post still makes sense, it was throat-clearing.
This tell is expensive on Instagram, where roughly the first 125 characters show before the caption collapses behind "more". A warm-up opener spends the entire visible caption saying nothing.
Tricolons and stacked adjectives
"Fast, simple and powerful." "Rich, bold, aromatic." Three is the most stable rhythm in English prose, so a model reaching for a shape reaches for three. The giveaway is that the third item is usually a synonym of the first.
Two rules kill it. Cap adjectives at one per noun, and require that adjective to be falsifiable. "Aromatic" cannot be checked by anyone. "Roasted eleven days ago" can. Second, any genuine set of three goes into a bulleted list where each line carries its own fact, so the rhythm has to earn itself. Then delete the third item and see whether anything was lost. If nothing was, it was there for the beat.
Generic sensory claims
"Buttery leather." "Melts on the tongue." "You will feel the difference." The model has never touched the product, so it produces the category average, and the category average is what your competitor's caption already says.
The fix is a supply-side rule: no sensory claim may appear unless it came from the brief. If the brief carries no sensory fact, the sentence becomes a structural one instead. Weight, origin, roast date, thread count, warranty, delivery window. These are duller to write and far better to read, because they are what a buyer is actually deciding on.
Family two: tells that dodge a decision
"Not only... but also"
Along with its cousins. "It's not just a bag, it's a statement." "More than a planner." The construction promises a contrast and then supplies two halves that mean the same thing. It is emphasis with no content.
Ban the strings themselves: "not only", "not just", "more than a". When a writer or a model hits the ban, one of two things happens. They find a real second fact, or they delete half the sentence. Both outcomes improve the post.
Hedging
"Can help." "May improve." "One of the most trusted." "Often leads to." Models hedge because they are trained to avoid overclaiming, and because nobody gave them a number they are permitted to state.
Do not fix this by deleting the hedge, which just converts a soft lie into a hard one. Fix it with a claims policy: a written list of what the brand may state outright, what needs a source attached, and what is never stated. Give the model the first list and it stops reaching for modal verbs, because it has something to say. Then ban "can help", "may", "could" and "one of the" in short copy. On X you have 280 characters, and every hedge costs you a fact.
Em dash overuse
The em dash is a universal joint. It attaches any clause to any other clause without requiring you to decide what the relationship between them is. That is exactly why models overproduce it, and exactly why it now reads as machine text.
Set a hard ceiling: zero in social copy, one per long-form piece. Every dash you remove forces a decision. The clause becomes its own sentence, or it takes a real conjunction, or it turns out not to be needed. Short sentences also read better on a phone, which is where all of this is being read.
Family three: the ask that asks for nothing
"Check it out." "Learn more." "What are your thoughts?" "Drop a comment below." The model knows a caption ends with a call to action, so it produces the shape of one. The shape contains no verb the reader can perform.
The rule: a CTA must name the action and say what happens after it. "Reply RESTOCK and we will message you when the 500g bag is listed" passes. "Comment your favourite" does not, because nothing follows. If you cannot say what happens next, delete the CTA. A post with no call to action beats a post with a fake one.
The constraint sheet
Every fix above, written as a rule someone can check. These belong in your stored brand rules, not in the prompt you retype each morning, because a rule you retype is a rule you will eventually forget.
| Tell | What it sounds like | Rule that removes it |
|---|---|---|
| Throat-clearing | "In today's fast-paced world..." | First sentence must carry a number, name, place, date or price |
| Tricolon | "Fast, simple and powerful" | Max two items in a prose run; three or more go in a list |
| Adjective stacking | "Rich, bold, aromatic blend" | One adjective per noun, and it must be falsifiable |
| "Not only... but also" | "Not just a bag, a statement" | Banned strings: not only, not just, more than a |
| Hedging | "Can help improve your reach" | No modal verbs in short copy; supply a claims policy instead |
| Em dash overuse | "Simple, and effective, and ours" | Zero in social copy, one per long-form piece |
| Generic sensory claims | "Melts in your mouth" | No sensory claim unless it appears in the brief |
| Empty CTA | "Check it out below" | CTA names one action and what follows it |
Before and after: one restock caption
A 4:5 feed post at 1080 x 1350 for a Bengaluru coffee brand. The brand is invented. The tells are not. This is what came back from a prompt that said "write an Instagram caption for our coffee restock" and supplied nothing else.
In today's fast-paced world, finding the perfect cup of coffee can be a challenge. Our rich, bold and aromatic blend is not only crafted with care, but also designed to elevate your mornings. Whether you are a seasoned coffee lover or just beginning your journey, this is one experience you will not want to miss. Check it out and let us know your thoughts below.
Same model, same product. This time the eight rules are stored against the brand and the brief carries nine facts.
The 500g Chikmagalur peaberry is back. Roasted Tuesday, 40 bags, shipping from Bengaluru on Thursday. ₹640, free shipping above ₹800. Grind coarse for the French press, fine for the South Indian filter. Reply RESTOCK and we will message you when the next roast is listed.
The second version is shorter and carries nine things a reader can check. The first carries none. Nobody asked the model to be more human. Somebody gave it something to say and removed the moves it uses when it has nothing.
Where the rules have to live
A rule pasted into a prompt survives one generation. A rule stored against the brand survives every generation, including the ones nobody watched. That is the entire argument for conditioning over prompting, and it is why "prompt engineering" as a daily habit keeps failing at volume.
In SocialPower that stored record is the Brand Brain: voice, audience, banned words, a hashtag ceiling and a claims policy, written once and applied to every caption, hashtag set and image prompt. It is not a paid extra. It conditions every step from brief to scheduled post on every plan, from Free at ₹0 to Business at ₹3,999 a month, all priced in rupees excluding 18% GST.
- 01Write the banned-words list first. Start with the strings in the table above, then add every phrase you have deleted twice from a draft.
- 02Set the hashtag ceiling. Instagram permits up to 30 hashtags. Choosing five forces the model to pick hashtags that describe this post rather than the category.
- 03Write the claims policy as three lists: state freely, state only with a source, never state. The first list is the one that stops the hedging.
- 04Put at least three facts the model cannot guess into every brief. Price, date, quantity, origin, dimension, or the name of the person who made the thing.
- 05Keep the approval gate on until the output stops surprising you. Autonomous Mode ships with that gate on by default and an enforced daily post cap, so the worst case is a queue you have to clear rather than a week of posts you would not have signed.
The two-minute pass before anything gets scheduled
- Read the first sentence on its own. Does it contain a fact? If not, delete it and start at the second.
- Count the adjectives. More than one per noun means the model was filling space.
- Find every dash. Replace each with a full stop and see whether anything actually breaks.
- Point at each claim and ask who could check it. Anything nobody can check is either a fact you still need to add or a line you need to cut.
- Read the CTA and try to obey it. If you cannot say what happens after you do, rewrite it or remove it.
- Compare word counts. A caption that lost 40% of its length in this pass was 40% filler, and it will be again next week unless a rule changes.
What rules cannot do
Constraints remove the tells. They do not supply the facts. A model does not know that Thursday's batch ran short, that four people in Mumbai asked the same question this week, or that the festive season order cut-off moved forward by two days. That information lives in your head and your inbox, and it is the only thing that makes a post yours rather than the category average with your logo on it.
So the workflow that holds up is unglamorous. You supply the facts, the model drafts and sizes them for each platform, and the stored rules stop it padding the gaps. Trend research and performance analysis stay manual work for now. Those agents are on the roadmap and labelled SOON, and saying otherwise would be exactly the kind of unfalsifiable claim this article is about.