There is a lot of confident advice about spotting AI writing, and most of it is wrong. Em dashes are not a tell. Neither is the word "delve". Plenty of people write that way, and plenty of AI output does not.
Here is what actually distinguishes it, and — more importantly — why you should be careful about concluding anything.
The real tells
Nothing is at stake
The clearest signal is not a phrase, it is an absence. AI writing very rarely commits to anything. It presents options, acknowledges complexity, and stops short of saying which one is right.
Human writing about something the writer cares about takes a position, and usually one you could argue with. Text where every paragraph could be agreed with by everyone is the strongest single indicator.
Perfectly even weighting
A model gives each section roughly the same depth, because it has no sense of which part is interesting. Humans over-write the bit they find compelling and rush the bit they find dull, and that unevenness is very hard to fake.
Specificity that goes nowhere
It will say "studies have shown" without a study, "many experts agree" without an expert, "significant improvements" without a number. Not lying exactly — reaching for the shape of evidence without the content.
Real writing about a real subject has grit in it: a date, a name, a number that could be checked, something that happened.
Structure over argument
Three neat points with three neat sub-points, each the same length, arriving at a conclusion that restates the introduction. It reads like a form being filled in, because in a sense it is.
No memory of itself
Longer AI text often repeats an idea in slightly different words several sections apart. A human writing an argument builds on what they already said; a model that has said something once is just as likely to say it again.
Why AI detectors do not work
They produce a probability from statistical properties of the text, and they are wrong in both directions often enough to be dangerous.
They clear real AI text. Ask the model to write in a plainer or more idiosyncratic voice and most detectors' confidence collapses. Anyone deliberately evading them succeeds easily.
They accuse real people. This is the serious failure. They flag non-native English speakers at markedly higher rates, because writing with simpler sentence structure and a smaller vocabulary looks statistically like generated text. They also flag people who write in a clean, plain style — which is what everyone is taught to do.
A tool that cannot distinguish "wrote carefully in a second language" from "used ChatGPT" should not be the basis of an accusation. Several universities have quietly stopped relying on them for exactly this reason.
A detector score is not evidence. It is a guess with a number attached, and the number makes it look like more than it is.
If you are accused and did not do it
- Ask what the evidence is. If the answer is a detector percentage, say plainly that these tools are known to be unreliable and to misclassify non-native speakers.
- Show your process. Drafts, version history, notes, a document's edit timeline. This is the strongest thing you can produce, and it is a good argument for always working somewhere that keeps history.
- Offer to discuss the content. Someone who wrote something can talk about it — why they chose that argument, what they cut, what they found difficult. Someone who generated it usually cannot.
If you are the one judging
Do not lead with a tool. Ask about the work. The gap between someone who wrote a thing and someone who did not shows up in about ninety seconds of conversation, and that conversation is fairer, more accurate, and harder to argue with than any percentage.
And be honest about what you are actually objecting to. Text edited for grammar by an AI, and text generated wholesale, are very different things that detectors treat identically.
---
Worth pairing with why AI gets things wrong, which explains where the confident-but-hollow quality comes from, and using AI to study without cheating yourself if the context is coursework.