Is Copyleaks Accurate? Reading Its Own Published Numbers

Copyleaks describes a 'dual-department' testing approach on its own site: a Data Science team and a separate QA team each test the detector against independent held-out datasets. In its Data Science team's test — 500,000 texts, split roughly 300,000 human and 200,000 AI-generated — Copyleaks reports a true positive rate of 0.988, a true negative rate of 0.999, and an F-beta score of 0.997. In its separately-run QA test of 248,555 texts, it reports 99.97% accuracy on human text (60 false positives out of 229,843) and 99.2% accuracy on AI text (148 false negatives out of 18,712).

Those are Copyleaks' own published figures, not the result of an outside audit — a caveat that applies to essentially every detector in this category, Copyleaks included, and worth keeping in mind before treating any single number as settled fact.

Copyleaks offers three detection sensitivity settings — Extra Safe, Balanced (the default), and Extra Sensitive — and publishes different false-positive and false-negative rates for each: Extra Safe at 0.009% false positives but 1.36% false negatives, Balanced at 0.026% false positives and 0.79% false negatives, and Extra Sensitive at 0.05% false positives and 0.53% false negatives, by the company's own figures.

The pattern is the usual trade-off in detection: turning sensitivity up catches more real AI text but risks flagging more genuine human writing, and turning it down does the opposite. Which setting is 'right' depends on whether a false accusation or a missed detection is the costlier mistake for the situation.

Copyleaks also publishes a claim specifically about non-native English writing — 99.84% combined accuracy and under 1% false positives across several academic writing datasets — again from its own testing, not an outside party. Separately, Copyleaks cites a handful of third-party academic studies as supporting its accuracy, but does so by summarizing them on its own blog rather than linking directly to fully independent, Copyleaks-specific test results, so treat those citations as a starting point for your own reading rather than a settled outside verdict.

More broadly, independent research on AI detectors as a category has found real limits worth knowing regardless of which tool you're using — accuracy dropping sharply once AI text has been paraphrased, and non-native English writing getting flagged at a higher rate than native writing. None of that is a claim about Copyleaks specifically outperforming or underperforming any other detector; it's a reason to treat any single score, from any detector, as one data point rather than a verdict.