As AI writing spreads, being able to judge whether something was likely machine-written is a genuinely useful skill — for teachers, editors, hirers and readers. There's no perfect test, and the automated 'AI detectors' are less reliable than they claim, but a trained eye can spot the common patterns.
This guide covers the honest tells of AI-written text, why detectors get it wrong, and how to make your own judgement.
What you need
- ✓The piece of writing in question.
- ✓An understanding of what human writing usually looks like.
- ✓Realistic expectations — this is judgement, not a certain verdict.
Step-by-step
- 1
Look at the tone and confidence
AI text is often smooth, evenly polished and confidently generic — it states things pleasantly without real insight, edge or opinion. Writing that never takes a position, never surprises, and reads a bit like a well-mannered brochure is a signal (though not proof).
- 2
Check sentence rhythm and structure
Models tend toward uniform sentence lengths, predictable paragraph shapes, and lots of tidy three-item lists. Human writing has more varied, sometimes messy rhythm. Very even, symmetrical structure throughout is a common tell.
- 3
Watch for stock phrases and filler
Repeated connectors ('moreover', 'in conclusion'), openers like 'in today's fast-paced world', and hollow phrases that add words without meaning are frequent in default AI output. A cluster of these raises suspicion.
- 4
Test the specifics
AI writing often stays abstract and can invent precise-sounding facts, quotes or citations that don't check out. Vague generalities, or 'facts' that fall apart when you verify them, are a stronger signal than any stylistic tell.
- 5
Don't rely on AI detectors
Automated detectors produce both false positives (flagging human writing, especially by non-native speakers) and false negatives (missing edited AI text). They can be a data point, but treating their score as proof is a mistake — people have been wrongly accused on them.
- 6
Judge in context
Consider what you know: the writer's usual style, the timeframe, whether specifics check out, and whether the piece shows genuine understanding or just fluent surface. Weigh the signals together rather than seizing on one.
Examples
- A student essay that's flawlessly smooth, evenly structured, full of 'furthermore' and generic claims, with a fabricated citation that doesn't exist — a strong pattern, confirmed by the bad source.
- An AI detector flagging a careful non-native writer's genuine essay as 'AI' — a classic false positive that shows why detectors can't be trusted alone.
Tips
- →Look for smooth, confident, generic prose with no real opinion or edge.
- →Uniform sentence rhythm and tidy three-item lists are common tells.
- →Verify specifics — invented facts and citations are the strongest signal.
- →Never rely on an AI detector's score alone; false positives are common.
- →Weigh signals in context rather than convicting on one.
Common mistakes
- Trusting AI detectors as proof. They have high error rates; use them as at most a weak data point.
- Judging on one tell. Weigh tone, structure, phrasing and specifics together, in context.
- Assuming polish means AI. Some people write cleanly; check specifics and voice, not just smoothness.
- Accusing without verifying. Confirm facts and consider context before concluding — false accusations do harm.
Conclusion
Spotting AI writing is judgement, not certainty: watch for smooth generic tone, even rhythm, stock phrases and thin specifics, and verify any facts. Don't trust AI detectors as proof — they're often wrong — and always weigh the signals in context before concluding.