How Accurate Are AI Detectors? False Positives and Misses
Short answer
AI detectors can make two kinds of mistake: flagging human writing as AI (a false positive) and missing AI text (a miss). Where the threshold sits trades one against the other, so a single accuracy figure says little. In our published tests, 96.9% of unedited 2026 AI text was labelled Likely AI, while 1% of human academic text was, and results get weaker on short texts, new models and rewritten text.
Key points
- There are two errors to ask about: false positives on human writing and misses on AI text. One accuracy number hides both.
- A threshold decides the trade-off. A stricter threshold protects honest writers but misses more AI text.
- Our detector is calibrated to keep false alarms low on human academic text: 1% and 0.7% on two held-out sets.
- Detection is weaker on short texts, rewritten or humanized text, the newest models and genres outside academic writing.
- No detector is perfect. Treat a score as an estimate and look for other evidence before drawing conclusions.
Two kinds of mistake
When people ask how accurate a detector is, they usually mean one number. But a detector can fail in two different ways, and they matter to different people.
A false positive labels human writing as AI. This hurts students, who may face an accusation for work they did honestly. A miss labels AI text as human. This matters to instructors who want to know when AI wrote a submission. A claim of "99% accurate" tells you nothing unless you know which of these it measures and on what kind of text.
Thresholds trade one error for the other
A detector produces a score, and a threshold turns the score into a label. Move the threshold up and fewer human texts are flagged, but more AI text slips through. Move it down and the reverse happens. There is no setting that removes both errors.
Because a false accusation is costly, we set our threshold to keep false alarms low on human academic writing. The scale has three verdicts, Likely human, Uncertain and Likely AI, and the Uncertain band exists so that borderline text is not forced into a confident label.
What our published numbers show
We calibrated and tested on 583 human academic texts and report the results by group. These are measurements of our detector on specific test sets, not promises about every text.
On human writing, the share labelled Likely AI was 1% and 0.7% on two held-out sets. By group it was 1% for student essays, 0.5% for preprint abstracts, 1.1% for papers by non-native English authors and 2.6% for encyclopedia text written before 2022. On AI text, 96.9% of unedited 2026 AI text was labelled Likely AI, and 96.5% on a public benchmark.
Where accuracy drops
Every detector has conditions under which it is less reliable, and it is more honest to list them than to hide them.
- Short texts: there is too little evidence in a few sentences. Our detector gives no verdict under 50 words.
- Newer models: detectors are trained on earlier output, so new systems can be harder to recognize until the detector is updated.
- Rewritten AI text: paraphrased AI text was still labelled Likely AI 84% of the time on a public benchmark, but only 50.5% when another AI model paraphrased it, and 59.5% after humanizer tools. Detection is weaker once text has been rewritten.
- Genres outside academic writing: on human general web text such as ads and press releases, 32.7% was labelled Likely AI, because the detector is calibrated for academic writing.
- Other languages: our detector is English only.
Why a false positive is not a rare freak event
A rate of about 1 in 100 sounds small, but across thousands of submissions it means real students will be flagged. That is why a detector result should start a conversation rather than end one, and why a flagged student deserves the chance to show their drafts and explain their work.
The reverse is also true. A clean result does not prove that no AI was used, since rewritten or lightly edited AI text can score lower.
How to use a detector sensibly
Used well, a detector is a screening tool. It points to passages worth a second look, which a writer can revise or an instructor can discuss.
Look at sentence-level highlights rather than only the headline number. Check the text length and genre. Compare against the writer's other work and ask about their process. You can try this free: AITurnitin checks up to 1,000 words without an account, and texts are deleted within 24 hours.
Frequently asked questions
What is a good accuracy for an AI detector?
Ask for both numbers: how often it flags human writing and how often it catches AI text, measured on realistic test sets. Ours labels 1% of human academic text and 96.9% of unedited 2026 AI text as Likely AI.
Can AI detectors be wrong about human writing?
Yes. Every detector produces some false positives, and they are more likely on formulaic text, very short samples and genres outside the detector's training. That is why a score should never be treated as proof.
Do detectors catch text that was paraphrased or humanized?
Partly. In our tests, 84% of paraphrased AI text on a public benchmark was still labelled Likely AI, 50.5% after another AI model paraphrased it, and 59.5% after humanizer tools. Detection is weaker, but not zero, and these tools can damage meaning and citations.
Are detectors getting better or worse as AI improves?
Both pressures exist. Models change, so detectors need regular updates and testing on new output. Any claim of permanent accuracy should be treated with caution.
Why do different detectors give different results on the same text?
They use different training data, scoring methods and thresholds, so a score from one tool is not comparable with another. Compare verdicts and highlighted sentences, not raw percentages.