How to Spot AI Writing as a Reader: A Field Guide to the Tells

Not a guide to fixing your own draft. This one is for the other side of the desk, editors, teachers, hiring managers, anyone reading text somebody else handed them and trying to work out how much of it a person actually wrote.

Most advice about recognizing AI writing stops at a list of overused words, delve into, unlock the power of, testament to. Those are real signals, and worth knowing, but they're also the easiest thing for a writer to strip out with one editing pass. The more durable tells live one level up, in structure and substance rather than word choice, and they're a lot harder to fake away because most people editing a draft never think to look for them, let alone fix them on purpose.

Structural tells

The predictable shape

Unedited AI output tends to follow the same macro-structure regardless of topic: an opening paragraph that restates the question being answered, three to five body sections of suspiciously similar length, and a closing paragraph that summarizes what was just said rather than adding anything new. A human writer working from genuine expertise tends to spend more space where the interesting material actually is and less where it doesn't, producing an uneven, lopsided structure. Perfectly even section lengths across a whole piece are a structural tell most people never consciously notice, even when they can feel that something's slightly off.

The reflexive summary

A bullet-point recap at the end, even when the piece was short enough that nothing needed recapping, shows up constantly in AI output and rarely in writing produced by someone with a genuine sense of when a reader needs a summary versus when it's just padding. If a five-paragraph piece ends with "key takeaways," that's worth noticing.

Content-level tells

Caveat stacking without a real position

Genuine expertise usually commits to a specific claim somewhere, even a hedged one. Text that qualifies every single statement, on one hand this, on the other hand that, with no sentence anywhere actually landing on a position, reads as evasive to a careful reader even when every individual sentence is reasonable. A piece with an actual point of view, stated plainly at least once, is a stronger signal of a human author with real experience than any phrase-level check.

Detail that's accurate but weirdly generic

This is the hardest tell to name and one of the most reliable once you notice it. AI-generated explanations tend to be correct at a general level and strangely empty of the specific, slightly odd detail that only comes from having actually done the thing being described. A real account of running a marathon mentions the specific mile where it went wrong. A generated one describes generic fatigue and generic determination. Accuracy without specificity is a pattern worth trusting more than any individual word choice, and it tends to show up across an entire piece rather than in just one sentence, which makes it more reliable than most tells on this list.

The triad habit

Watch for things consistently arriving in groups of three, three adjectives, three examples, three reasons, applied so consistently it starts to feel like a tic rather than a genuine rhetorical choice. Occasional triads are completely normal in good writing. A piece where nearly everything comes in a tidy set of three is worth a second look.

Reading for the combination, not any single tell

Any one of the patterns above shows up sometimes in writing that's entirely human, which is exactly why relying on a single tell produces so many false accusations. A careful writer occasionally hedges everything in one paragraph. A tired one occasionally reaches for a generic phrase. No individual pattern is worth trusting much on its own. The real signal is several of them stacking up in the same piece: predictable structure and caveat-stacking and suspiciously even section lengths, all present together rather than one flag in isolation.

This is closer to how a doctor reads a set of symptoms than how a spell checker flags a single typo. One symptom rarely means much on its own. A cluster of them pointing the same direction is worth taking seriously, though even then it's a strong hypothesis rather than a confirmed diagnosis. Treating a pattern match as proof is where this kind of reading goes wrong most often, and it's worth resisting that jump even when the pattern feels obvious.

What isn't a reliable tell

Its worth being just as clear about the false positives, because chasing the wrong signal causes real harm, wrongly accusing a student or a colleague of using AI when they didn't. A few things that feel like tells but aren't reliable on their own:

• Clean grammar. Careful human writers produce clean grammar too. Its evidence of care, not evidence of authorship.
• Short, simple sentences. Some of the best human writers in the world favor short sentences. This is a style choice, not a tell.
• Unfamiliar vocabulary or phrasing patterns. This one causes real damage. Non-native English speakers and people writing in a second language get flagged constantly by both human readers and automated detectors for patterns that are simply normal features of their actual writing background, not AI involvement.
• A single instance of a common AI phrase. Everyone occasionally writes "it's important to note." One instance is noise. A cluster of several tells together, structural and content-level, is signal.

The cost of getting this wrong runs in both directions and neither one is small. Wrongly clearing actual AI-generated work as human misses exactly the thing the check was meant to catch. Wrongly flagging a real person's honest writing, especially someone whose natural style happens to overlap with these patterns, causes direct harm to someone who did nothing wrong. Neither mistake is more acceptable than the other, which is the real argument for treating any of this as a starting point for a closer look rather than a conclusion by itself.

None of the patterns in this piece amount to proof on their own, and stacking a few weak signals together still isn't certainty. Reading for tells isn't the most rigorous option available right now, either. A cryptographic watermark check, when one is available, is a fundamentally different and stronger kind of evidence than any pattern a human reader can notice. Worth knowing that distinction if a real decision, a grade, a hire, a byline, actually depends on the answer.

Reading, not accusing

The point of learning these patterns isn't to turn every piece of writing into a suspect lineup. Its to read a little more carefully, and to know the difference between a hunch worth investigating further and a hunch that's really just an unfamiliar style getting mistaken for something it isn't. Structure and substance are more durable signals than any single phrase, but even taken together they're still a pattern match, not a verdict. Treat them that way and this becomes an actually useful reading skill rather than a tool for jumping to conclusions about someone else's work.

Where this actually matters most is in the moment right before a real decision gets made on the back of a hunch, before a grade gets lowered, before a byline gets questioned, before a submission gets rejected without a second look. That's the point to slow down and ask whether the evidence is a genuine cluster of the patterns above or just one unfamiliar sentence that happened to stand out. The habit worth building isn't suspicion by default. Its noticing enough to ask a better question before acting on a first impression, and treating the answer to that question with the same caution any pattern match deserves.