Can People Tell If Content Is Written by AI?

Mostly, no. In controlled tests, readers identify AI-written text with roughly the accuracy of a coin flip. What readers do have is a confident, shared idea of what AI writing looks like — and a 2023 study mapped that idea in detail. It turns out to describe, almost point for point, what makes any content feel generic. That difference decides whether your posts read as yours or as anyone's.

Where we went looking

In early 2023, the journal PNAS published a study titled "Human heuristics for AI-generated language are flawed", by Maurice Jakesch, Jeffrey Hancock and Mor Naaman. The setup sits unusually close to a business owner's daily work. Across several experiments, thousands of participants read short self-presentation texts — dating profiles, professional bios, Airbnb host profiles. Some were written by people, some by a language model. The task fit in one line: spot the machine.

Self-presentation is the genre you work in every time you post. A caption about your bakery, your practice or your workshop is a short text whose job is to make a stranger trust you. If readers could reliably spot AI in that genre, the fear that clients "will just know" would be justified. So we read the paper closely.

How accurate were readers at spotting AI?

Close to chance. Across the three domains, participants picked the right author about half the time — the experiments report accuracy barely above 50 percent, the score of blind guessing. Practice barely helped. And the domain didn't change the result: people judging host profiles, the format closest to a small business presenting itself, did no better than the rest.

One detail matters more than the headline number. Participants were not hesitant. Their votes clustered on the same texts — they broadly agreed about which ones "felt AI". They shared a mental model of machine writing. It just didn't match how machines write.

“

Readers don't detect AI. They detect writing that could sit under any logo.

”

Which cues do readers actually use?

The researchers isolated the signals behind those judgments. Texts using first-person pronouns, contractions, or mentions of family and personal life were judged human. Texts using long or rare words were judged AI. Spontaneity read as human; polish read as machine.

Every one of those cues fails. Language models produce "I", "we" and contractions effortlessly, and they mention personal life whenever the prompt points there. Many professionals, meanwhile, write formally — and get classed as machines for it. The authors pushed one step further: because the cues are stable and predictable, text can be tuned against them. They generated profiles that participants rated as more human than the ones real people had written. The paper calls this "more human than human".

So what are readers reacting to in your feed?

Interchangeability. Strip the flawed cues away and one signal keeps working for readers: whether a text could have been written by anyone. The study's participants called texts "AI" when nothing in them was anchored to a specific person. As a test of authorship, that reflex misfires. As a test of writing, it is fairly sound.

This matches a pattern we see in operator feedback on generated captions. The captions business owners flag as "sounds like AI" share surface habits, and none of them concern the generator. The sentence that addresses two audiences at once instead of one. The opener that announces a theme before saying anything. The engagement question bolted onto the end. Meanwhile, captions that carry one concrete trace of the business — a material, a street name, a season, a service named the way the owner names it — almost never get flagged. Whoever wrote them.

The mechanism is plain. Readers judge writing the way the study's participants did: by how interchangeable it feels. A sentence that could sit under any logo reads as machine. A sentence only your business could have produced reads as human. The generator is invisible. What shows is the input.

What should you change in your posts?

Four moves, all doable this week.

Keep first person and contractions. They read as human, they are natural in a caption, and dropping them costs you the one bias working in your favor.

Replace category language with your own vocabulary. "Quality craftsmanship" and "your trusted partner" are the average of every business in your category — the exact texture readers flag. The words on your website, the way you name your services, the phrases clients repeat back to you: that is the material that makes a post yours. An AI that has read your site writes with those markers; an AI given a one-line description writes the category average. The input, not the tool, sets the voice your posts carry across your social networks.

Cut the three flagged habits: the two-audience address, the theme-announcing opener, the decorative closing question. None of them add information; all of them add typicality.

Put one untransferable detail in each post. A detail that would be false under a competitor's logo — the street, the supplier, the season's first delivery, the price of the thing you actually sell. One is enough to move a caption from "anyone" to "you".

The goal is not passing as human. The goal is being specific — which, conveniently, is also what the study says readers reward.