Are AI detectors accurate?

The short answer

No, and not in a way that gets better with a newer detector. An AI text detector isn't finding a hidden mark, because there isn't one to find. It reads your writing, estimates how predictable the word choices look, and turns that guess into a confident-looking percentage.

OpenAI ran its own detector for six months in 2023 and then withdrew it for low accuracy. Its published numbers: it caught 26% of AI-written text, and wrongly flagged 9% of human writing. The company that built the model couldn't reliably spot the model.

Who gets caught

A 2023 Stanford study ran several detectors over TOEFL essays by non-native English speakers. More than half were flagged as AI-generated. Essays by native-speaking US school students were classified almost perfectly.

The reason is mechanical, not malicious. Simpler vocabulary and steadier sentence rhythm look like a machine to a tool that measures predictability. So the people least equipped to argue back are the ones most often accused.

What "92% AI" actually means

It isn't 92% certainty. It's a score from a model that never saw your drafts, your sources, or you. You can't reproduce it, and nobody can tell you which sentences caused it. A number you cannot interrogate is not evidence, however precise it looks.

What you can check instead

Narrower things, which are real. Hidden Unicode characters in pasted text are countable and visible — run any text through the checker and you'll see exactly which characters are there and where. Images can carry content credentials that are cryptographically signed, so tampering breaks them.

Both tell you something specific and verifiable about a file. Neither tells you who wrote it, and nothing currently available does.

If you have to make a decision about a person, ask for the draft history and have a conversation. Don't outsource it to a percentage.