Are AI detectors accurate?

The short answer: accurate enough to be useful, not accurate enough to be proof. AI detectors catch a lot of unedited AI text. They also flag human writing, miss edited AI writing, and disagree with each other.

What independent tests found

In 2023 a group of academic integrity researchers tested 14 detection tools on human-written, AI-written and edited text. Their conclusion was that the tools were neither accurate nor reliable. Every tool scored below 80% accuracy and only five scored above 70%.

Two patterns stood out. The tools leaned toward calling a text human, so a good share of AI text went through unflagged. And once AI text had been paraphrased or edited, performance dropped sharply.

Detectors have improved since then. So have the models they are trying to catch.

False positives: human writing flagged as AI

The costlier mistake runs the other way. A 2023 study from Stanford ran 91 essays written by non-native English students through seven detectors. On average the detectors flagged about 61% of them as AI. Nearly all of the essays, 97.8%, were flagged by at least one detector. Essays by American eighth-graders were almost never flagged.

The reason is uncomfortable. People writing in a second language tend to use a smaller vocabulary and simpler sentence structures. To a detector, simpler and more predictable looks like a machine.

Even detector makers have been cautious

In January 2023 a leading AI lab released its own detector. By its own figures, it correctly identified 26% of AI-written text and wrongly labeled 9% of human writing as AI. The lab withdrew the tool that July, citing its low rate of accuracy.

When detectors are most reliable

1. 1Longer texts. A few hundred words give a detector far more to measure than a few sentences.
2. 2Unedited output. Text pasted straight from an AI tool is the easiest case.
3. 3Ordinary prose in English, which is what most detectors were trained on.

When they are least reliable

1. 1Short texts.
2. 2AI drafts that a person has edited heavily, and human drafts that AI has polished.
3. 3Translations.
4. 4Very formal or technical writing, where human authors are also even and predictable.
5. 5Writing by people working in a second language.

How to read a score

Treat it as evidence, not as a verdict. Look at which sentences were flagged and whether they have anything in common. If the stakes are high, run the text through more than one detector and ask for what a detector cannot see: drafts, notes, version history, a conversation about the work.

No decision about a person should rest on one number from one tool.

If your own writing was flagged

1. 1Keep your drafts and version history. They are the best evidence you have. Our guide on what to do if you are falsely accused covers the rest.
2. 2Look at the flagged sentences in a tool that shows them, such as the neepora AI detector.
3. 3Rewrite those sentences yourself. Vary the length, drop the stock phrases, add a detail only you would know. Why writing gets flagged as AI lists the nine usual causes.
4. 4Talk to the person who raised it, and bring the drafts.

Sources

Weber-Wulff et al., Testing of Detection Tools for AI-Generated Text (2023).

Liang et al., on detector bias against non-native English writers (Patterns, 2023).

Announcement and withdrawal notice for an AI text classifier (2023).

Keep reading

Why is my writing flagged as AI?

Falsely accused of using AI? What to do

How AI detectors work

Check a text with the AI detector

Updated October 7, 2026