Do AI detectors work?

Three studies carry most of the weight, and each measured something different.

Weber-Wulff and colleagues built 54 documents in six classes, from pure human writing to AI text paraphrased by a tool, and concluded that the detectors were "neither accurate nor reliable" and leaned toward classifying text as human; even the highest accuracy values were below 80 percent (Weber-Wulff et al. 2023, International Journal for Educational Integrity, read 9 November 2026). Liang and colleagues found that seven detectors misclassified over half of 91 human-written TOEFL essays, with an average false positive rate of 61.22 percent, and that 89 of the 91 were flagged by at least one detector (Liang et al. 2023, Patterns00130-7), read 9 November 2026). OpenAI's classifier caught 26 percent of AI text and mislabelled human text 9 percent of the time, and OpenAI withdrew it on 20 July 2023 (OpenAI, new AI classifier announcement, read 9 November 2026).

Three dated findings, all from 2023: detectors disagreed, missed edited text and flagged some human writers far more than others.

The vendors answer with their own numbers. Every figure below is the vendor's, measured on the vendor's own data.

Turnitin states it keeps its false positive rate "under 1% for documents with over 20% of AI writing", validates that figure on over 700,000 academic papers written before ChatGPT, and accepts that it may miss some AI text to hold the rate (Turnitin, AI writing detection FAQ, read 9 November 2026). GPTZero states that its confidence categories are tuned so the average error rate is under 1 percent for high-confidence predictions and that its de-biasing work has cut the false positive rate on TOEFL texts to 1.1 percent (GPTZero, technology page, read 9 November 2026). Pangram Labs' technical report states its classifier has over 38 times lower error rates than leading commercial tools on a benchmark of 10 domains and 8 models, and that it is not biased against non-native speakers (Emi and Spero 2024, arXiv 2402.14873, read 9 November 2026).

None of these has been reproduced in a peer-reviewed study this site could find. That is not an accusation; it is the reason every reading on this site carries the detector's name and the date.

The failure modes are consistent across the studies and the vendors' own limitation pages.

Two of these deserve a line each. The non-native finding means a detector can penalise a writer for limited vocabulary, which is a fairness problem before it is an accuracy problem; the guide on AI detector false positives covers what that looks like for the person flagged. The edited-text finding is the reason humanizers exist, and the guide on whether AI humanizers work sets out what has changed since 2023.

The vendors are clearer about this than their customers sometimes are. Turnitin states the percentage "should not be used as the sole basis for action or a definitive grading measure by instructors" (Turnitin, AI writing detection FAQ, read 9 November 2026). GPTZero's FAQ states that "these results should not be used to punish students" and recommends a holistic assessment (GPTZero, FAQ, read 9 November 2026). OpenAI wrote that its classifier "should not be used as a primary decision-making tool" (OpenAI, new AI classifier, read 9 November 2026).

No detector score should be the only evidence in a misconduct case.

A result is a probability from one model on one date. Used that way, as one input beside drafts, version history and a conversation with the writer, detectors are informative. Used as a verdict, they fail the people the studies say they fail most.

The published method runs every humanizer output through five detectors, GPTZero, Originality.ai, Copyleaks, Pangram and ZeroGPT, and records each one's own label and the date. Five because they disagree, and the spread is the finding. Turnitin is not among them because Turnitin's detector is not available to the public; this site reports Turnitin's published statements and never a Turnitin result for any tool.

Every input is also checked on every detector before humanizing, so the change a tool makes is visible against a baseline, and six human-written control texts show what each detector's false positive behaviour looks like on the same genres. The guides hub holds the rest of this series, including what detectors measure and what to do if you are flagged.