Is This Written by AI? How to Spot Machine-Written Text

The email is polished. The cover letter is flawless. The review is glowing. And you can't shake the question every reader now carries: did a human write this?

Spotting AI text honestly is harder than spotting AI images — good human writers and well-prompted models overlap heavily. But unedited machine text still has a recognisable accent. Here's what actually gives it away, what doesn't, and why the detector tools you've heard about aren't the answer.

First, Kill the Myths

The folk tells are mostly wrong. Em-dashes, the word "delve", semicolons, long words — humans have used all of these for centuries, and accusing every literate writer of being a robot is the current era's silliest reflex. One quirky word proves nothing. Patterns prove things.

The Real Tells

• Fluency without friction. Every sentence lands smoothly, and none of them surprises you. Human writing has snags — an odd phrase, a too-short sentence, a joke that half-works. Unedited AI text is a river with no rocks.
• Structure addiction. An intro that previews, three tidy points, a conclusion that recaps, everything in parallel form — for a message that needed two sentences. The model's essay reflex applied to a lunch invitation.
• Confident vagueness. Sentences that sound informative and commit to nothing: "This approach offers numerous benefits across various contexts." Ask what it actually said, and the answer is fog.
• Hedged symmetry. "While X has advantages, it's important to consider Y." Both sides, every time, offending no one. Humans have opinions; default AI has balance.
• No lived detail. Real writers leak specifics — the client's weird deadline, the rain that day, the exact error message. AI text describes categories of experience without ever having had one.
• Uniform register. Humans drift between formal and casual mid-message. Machine text holds one temperature from first word to last.

One of these means nothing. Four of them in the same piece is a strong signal.

Why Detector Tools Disappoint

AI-text detectors misfire in both directions: they flag non-native English speakers and formal human writing as machine-made (with real consequences for students), and they miss any AI text that's been lightly edited. Treat their verdicts as one weak signal, never as evidence for an accusation. Your calibrated judgement, applied to the patterns above, genuinely outperforms them — the case we make in Everyone Needs an AI Radar Now.

The Question Behind the Question

Sometimes "is this AI?" is the wrong question. A colleague using AI to tidy a report they'd drafted is workflow, not deception. What matters is whether the facts are checked and the sender stands behind it. Save your suspicion for where authorship is the substance: reviews, testimonials, "personal" outreach, student work, and anything asking for money or trust.

(Understanding why models write this way — the next-word guessing under the hood — makes the tells click into place. Ten fun minutes: How LLMs Actually Work.)

Calibrate Your Eye, Daily

Like image-spotting, text-spotting is a trainable skill that most people overestimate until they're tested. Our free daily game AI or Not includes human-vs-machine writing rounds alongside its real-vs-AI photo rounds — 8 rounds a day, instant reveals, streaks to keep you honest.

Play a week and you'll notice the shift: from "this feels off" to "structure addiction, confident vagueness, no lived detail — machine." Start with today's rounds, and see the image-side tells in How to Tell If an Image Is AI-Generated.