How to Spot AI-Written Text (and Why Detectors Miss)

There are real tells that a piece of writing came from an AI, but there is no reliable way to be sure, and the tools that claim to measure it are not the answer. Careless machine text has habits you can learn to spot. Edited machine text hides them. And the automated detectors that promise a verdict are wrong often enough to end careers and fail innocent students. So learn the patterns, hold them loosely, and never treat a hunch or a detector score as proof.

What actually gives AI writing away?

A cluster of habits, not a single fingerprint. Left on autopilot, a model tends to write in a certain way, and once you have seen the pattern a few times it gets easier to notice. Here is what to watch for, with the honest caveat that every one of these also shows up in real human writing.

Over-hedging and stacked qualifiers.Models love to soften. You get sentences that pile up “generally,” “often,” “in many cases,” and “it depends” until the claim underneath has no edges left. A confident human writer usually just says the thing.

Suspiciously even structure. AI output often comes out weirdly symmetrical. Three neat paragraphs of nearly equal length, each opening the same way, every list a tidy set of parallel items. Real writing is lumpier. It runs long where the idea is rich and short where it is not.

Generic examples with no lived specifics.Ask a model for an example and you often get a placeholder: “a small business owner,” “a busy professional,” a Sarah who runs a bakery. Real experience leaves fingerprints. The odd brand of coffee, the exact error message, the Tuesday it went wrong. Vague, frictionless examples are a quiet giveaway.

Uniform sentence rhythm. Human prose has burstiness. Long sentences bump against short ones. Fragments happen. Machine text tends toward one comfortable length, sentence after sentence, and the evenness starts to feel like a metronome once you notice it.

Punctuation and filler habits.Heavy, decorative use of the em-dash is a known lean, as is filler like “it’s important to note,” “in today’s world,” and neat closers that restate what was just said. None of these is damning by itself. A careful human might use any of them.

Tidy but flavourless. This is the hardest one to name and often the most telling. The grammar is clean, the paragraphs flow, and yet nothing surprises you. No strong opinion, no risk, no voice. Competent, and forgettable.

The trap is treating any of these as a verdict. A plain, careful, well-organised human writer can hit every marker on this list. That is exactly why the confident-looking tools built to catch AI fall apart, which is where this gets serious.

Why don’t AI detectors work?

Because they guess at a pattern rather than see the truth, and they guess wrong far too often to be trusted with a decision. An AI detector reads statistical features of the text, roughly how predictable each word is, and returns a percentage that looks authoritative. It is not.

Start with false positives. Independent testing through 2025 and into 2026 keeps finding that detectors flag genuine human writing as AI at rates high enough to matter. The people hit hardest are non-native English speakers. A widely cited study found detectors wrongly flagged a large share of essays by non-native writers while barely touching essays by native speakers, because simpler, more predictable phrasing is exactly what these tools read as machine-made. The bias is baked into how they work.

Then there is how easily they break. A few minutes of editing, a paraphrase, a swap of a few words, and detector accuracy drops sharply. The same tools that catch raw, unedited model output miss it once a human has tidied it up, which is precisely the case you would most want to catch. Some vendors even publish wide error margins on their own scores, so a result reading “50% AI” might really sit anywhere across a broad band.

The deepest problem is that there is no ground truth. The detector never knows who wrote the text. It infers, and an inference dressed up as a percentage invites people to act as if it were fact. By 2026 that pressure had pushed some universities to switch AI detection off entirely over reliability and bias concerns, and the major vendors state plainly in their own documentation that a score should not be the sole basis for an accusation.

So take the strong version of the advice. Do not use detector scores to set grades, screen hires, or accuse anyone of cheating. The cost of a false positive is somebody’s reputation or livelihood, and the tool cannot carry that weight. If you want the wider case for treating AI claims with this kind of care, the the AI literacy guide lays out the habits of mind that hold up.

What should you check instead?

Check the things a machine cannot fake: facts, sources, and specifics. Authorship is usually the wrong question. What you actually care about is whether the text is true, careful, and accountable, and you can test that directly.

Verify the facts and the sources. Models invent confident, wrong details, including citations and quotes that do not exist. Pull one or two claims and check them. Click the sources. If a referenced study or figure cannot be found, that tells you far more than any detector, whoever or whatever wrote the page.

Look for checkable specifics. Real knowledge tends to be concrete: a named tool, an exact number, a date, a detail that could be proven wrong. Writing that stays abstract and hedged whenever it should get specific is a sign of thin understanding, human or machine.

Consider context and provenance. Where did this come from? A named author with a track record, a site that stands behind its work, an email you can reply to. Provenance will not tell you the exact tool used, but it tells you who is accountable, which is the thing that usually matters.

Ask the author. The single most reliable test is a conversation. Someone who did the work can explain the choices, the dead ends, the reason they framed it this way and not that. Someone who pasted a prompt usually cannot. In a classroom or a hiring process, that beats any score.

It helps to understand the other side too. The better you get at writing clear instructions, the more you see how model text takes its shape. A short read on how to write better prompts shows you what the machine is responding to, which sharpens your eye for its defaults.

The bigger picture

Text is only half of it. The same story is playing out with pictures and video, where the tells are different but the lesson is the same: the giveaways are fading, and the reliable move is to verify rather than eyeball. The guide to how to spot AI images and deepfakes walks through the visual side.

The real skill here is not memorising a checklist. It is building the instinct to slow down, notice when writing feels frictionless, and go check something before you decide. That instinct improves with reps. You can train it on real examples in the Spot the AI text practice tool, where you guess and then find out, which is a faster way to calibrate than any rule you could read.