ChatGPT vs Human Writing: Can You Tell the Difference?

Most people are confident they can spot ChatGPT. Most people are wrong about how often they can. Research on this keeps finding the same thing: readers asked to separate human from AI passages do only a little better than a coin toss, and experienced readers aren't dramatically better than everyone else.

So what are the real differences? Not the ones people usually point to.

It isn't the vocabulary (mostly)

Yes, certain words became running jokes. 'Delve'. 'Tapestry'. 'Testament to'. But models update, users edit, and one unusual word proves nothing. People write 'delve' too.

It's the shape

The more reliable tells are structural. AI prose tends to be even: sentences of similar length, paragraphs of similar length, points delivered in threes, and a neat summary at the end of each section. It qualifies everything, rarely commits to a view, and almost never includes a specific detail that couldn't have come from the prompt.

Human writing is messier. People digress. They put the most important point in a four-word sentence. They reference something that happened last Tuesday. They're sometimes wrong in interesting ways.

Why detectors do better than people

Detectors don't read for meaning; they measure predictability. That makes them better than humans at flagging raw model output, and worse at almost everything else: short texts, heavily edited drafts, formulaic human writing, and writing by non-native speakers, which they flag far too often.

Does it matter?

Depends on the job. For a product description, nobody cares who typed it as long as it's accurate. For an essay, an application or anything where your own voice is the point, it matters a lot, and not mainly because of detectors. Readers trust writing that sounds like a person who knows something. Generic text loses them, whoever wrote it.

That's the case for editing AI drafts properly: add the specifics, break the pattern, and cut anything you wouldn't actually say.