How accurate are AI detectors? Every study, indexed

How accurate AI detectors are depends on who measured, on what text, and when. This hub holds one record for each published evaluation the index has read in full: fifteen independent studies, three vendor or lab reports and one human-rater study, 35 records in all as of 9 November 2026, nine of them on the morning of that day, eight added later the same day (Sadasivan 2023, Krishna 2023, Orenstrakh 2023, Gao 2023, Jabarian and Imas 2025, Van Vlasselaer 2026, NIST AI 700-1 and Ha and Passananti 2024) and two in the fifth wave (Pegoraro 2023, a 20-method sweep on ChatGPT question-and-answer text, and Fleckenstein 2024, in which 89 novice and 200 experienced teachers judged essays and no software was tested). Each record states who ran the test and when, which detectors it covered, the sample by class and count, the results exactly as the paper reports them with their base, the limits the authors state, and the cards in the index the record feeds. Nothing on a record comes from a write-up of the paper; every figure was read in the paper itself. The methodology page sets out what a study needs before it is indexed.

The thirty-five records do not agree, and the hub does not average them. Weber-Wulff and colleagues tested fourteen detectors on 54 documents in spring 2023 and found that no tool reached 80% on their averaged accuracy measures, with Turnitin highest at 41 of 54 documents classified correctly (DOI 10.1007/s40979-023-00146-z). Walters ran sixteen detectors over 126 student and generated essays a few months later and found Copyleaks and Turnitin correct on all of them, with other tools far lower on GPT-4 text (DOI 10.1515/opis-2022-0158). Liang and colleagues reported an average false positive rate of 61.22% across seven detectors on 91 essays by non-native English writers, against near-perfect accuracy on essays by US eighth graders (DOI 10.1016/j.patter.2023.100779). Jabarian and Imas 2025 ran four detectors over 1,992 human and 7,968 AI passages and report false positive rates at a 0.5 threshold of 0.0000 for Pangram, 0.0011 for Originality.ai, 0.0071 for GPTZero and 0.9503 for the RoBERTa-base baseline (NBER w34223, read 9 November 2026); Van Vlasselaer 2026 ran four tools over 160 academic papers and found Turnitin, GPTZero and Copyleaks scoring 70% to 100% of the fully AI-generated papers under 20% while all four cleared every human paper (DOI 10.1007/s40979-026-00226-w, read 9 November 2026). The same tool can look strong in one record and weak in the next because the samples, versions, thresholds and metrics differ. That is why every value on a record carries its base, and why the rankings derived from these records count evidence and report rates rather than declare a winner.

The table below is the hub's view of the thirty-five records: register ID, the study's short name, the year, how many detectors it tested, the size of its sample as the record states it, and the metric it reports. The three vendor or lab reports, the OpenAI classifier announcement of 2023, Turnitin's statements on its own detector and the NIST GenAI pilot report of 2025, are indexed in their own column of the table as reports rather than studies, so that a reader can weigh them apart from the fourteen independent evaluations. One record, Ha and Passananti 2024, scored images rather than text, and the ranking by content type reads it.

Records feed cards. When a study tested a tool that has a card in the detector index, the card's Independent studies field counts the record and quotes its result with a link back here. 35 of the 72 cards have at least one record; GPTZero has twelve and the GPT-2 Output Detector ten, both lifted by one when Pegoraro 2023 was added on 9 November 2026. A card with no record says so by leaving the field out, and the ranking of independently tested detectors lists only the cards that have one. Where a record reports a false positive rate on human-written text, that figure also enters the false positive rate ranking, again with its base and the study year; where it reports a miss rate or a result after a paraphrase, the false negative rate ranking and the robustness ranking read it, and the 2026 false positive evidence review lists every false positive figure in one table.

On 10 November 2026 seven records gained a card row without a new record being added: Weber-Wulff 2023 now feeds PlagiarismCheck, Walters 2023, Kar 2024 and Alshammari and Rao 2025 feed Content Detector AI, Orenstrakh 2023 and Perkins 2024 feed GPTKit, and the RAID benchmark feeds Fast-DetectGPT; the Q2 2027 studies digest lists the seven with the figure each new card records.

Later the same day the tenth wave gave five records a further card row: Weber-Wulff 2023 now feeds Check For AI and Writefull (12 of its 14 tools carded), Pegoraro 2023 feeds Writefull, Kar 2024 feeds Wordtune, and the RAID benchmark and DAMAGE 2025 feed Binoculars, whose 0.00% at threshold 0.5 and 80.3 under paraphrase enter the false positive and robustness rankings; the four non-text cards of the same wave attach to no record, and the three voice and video challenges still name no indexed product.

The eleventh wave, the same day, added five records and gave four older records new card rows: Rashidi 2023 ran the RoBERTa GPT-2 detector on 14,400 real journal abstracts and feeds the GPT-2 Output Detector card's base (up to 8.69% of a 1,600-abstract range above a 50% score, no AI text scored); ASVspoof 2021 and ADD 2023 are voice challenges and FaceForensics++ 2019 a video benchmark, none scoring an indexed card; and the Originality.ai compilation is the vendor's own meta-analysis of 17 studies, recorded as a vendor statement on the Originality.ai card and never as a study result. Walters 2023 now feeds GPT Radar, Grammica, IvyPanda and SEO.ai (14 of its 16 tools carded), Liang 2023 feeds AI Writing Check, and the RAID benchmark and DAMAGE 2025 feed RADAR; the launch year ranking dates each of these tools from its own page.

A study describes the versions it tested on the dates it tested them. Detectors change after publication, and several vendors have replaced their models since 2023, so a 2023 result is a dated fact about a dated version, not a statement about today's tool. Each record shows its year beside every figure for that reason, and the home board shows which AI detectors exist and who has tested them in one view. To suggest a study for the index, send the citation through the contact page; accepted additions are logged in the corrections log with their date and gathered in the dated index updates.

No record joined in the twelfth wave on 10 November 2026; the Q3 2027 studies digest lists the five eleventh-wave records with their figures and bases, and the 2027 census tabulates all 35 records by year and kind, 17 of them from 2023.