The research.
A register of published research on how reliably AI-writing detectors read human-written text. Each entry states what a study tested and what its authors reported, in the study's own figures, dated and linked to the original. The studies measure particular tools on particular datasets and do not yield one aggregate error rate; this page reports them and determines nothing about any case. If an outcome could affect enrollment, immigration status, a degree, professional licensing, or accommodations, consider obtaining qualified advice before responding.
What the published studies report
Newest first. Preprints are labeled as preprints; a vendor's statement about its own product is labeled as a vendor statement. A finding about a measured population is stated at that population and not extended past it.
- Thirteen detectors tested on edited non-native academic writing — preprint, August 2026
Park, Jeong and Kim, “Style as a Confound: False Positives in AI Detection of Non-Native Academic Writing” (arXiv, submitted Aug 27, 2026; presented at EMNLP 2026), examined 135,389 document pairs from a professional English editing service (2018–2025): manuscripts by non-native English writers and their native-edited versions, so that authorship and content were held constant while style changed. Across 13 publicly available detectors, false-positive rates on the human-written texts ranged from 0.0% to 100.0%. The same edits raised AI scores in some detectors and lowered them in others, and score changes correlated with the extent of editing. The authors identify professional editing style as a confounding variable in detector output. The study did not test proprietary commercial detectors, and its documents are academic manuscripts, not student essays. arXiv:2608.26710. Retrieved Oct 14, 2026.
- Keystroke-timing verification tested against forgery — preprint, January 2026
“On the Insecurity of Keystroke-Based AI Authorship Detection” (arXiv, submitted Jan 24, 2026) examined proposals to verify human authorship from typing rhythm rather than from the text itself. Testing five classifiers on 13,000 typing sessions from the SBU corpus, the authors reported that 99.8% or more of forged or transcribed sessions were classified as human under the attacks they specified — including the case of a person simply copy-typing machine-written text — while fully automated injection remained easy to catch. The authors present this as a security failure of timing signals used alone; the study did not evaluate every process-tracking design. arXiv:2601.17280. Retrieved Sep 20, 2026.
- Sixteen detectors tested for demographic bias — preprint, December 2025, revised April 2026
Stowe et al., “Identifying Bias in Machine-generated Text Detection” (arXiv), assessed sixteen detection systems on a curated dataset of student essays across four attributes: gender, race/ethnicity, English-language-learner status, and economic status. The authors reported that biases were generally inconsistent across systems, but that essays by English-language learners were more likely to be classified as machine-generated, with non-White ELL writers disproportionately flagged relative to White ELL writers, while essays by economically disadvantaged students were less likely to be flagged. Human readers annotating the same essays performed poorly at detection overall but showed no significant biases on the studied attributes. arXiv:2512.09292. Retrieved Sep 20, 2026.
- A national survey of who reports being flagged — Common Sense Media, 2024
A survey, not a detector measurement, entered at its own scope: in “The Dawn of the AI Era” (2024), a national survey of teens and parents, 20% of Black teens reported that a teacher had wrongly flagged their schoolwork as AI-generated, against 10% of Latino teens and 7% of white teens — a K–12 finding about reported experience. Common Sense Media, 2024. Retrieved Oct 14, 2026.
- Purpose-built detectors tested for bias on GRE essays — Jiang et al., 2024
Jiang, Hao, Fauss and Li, “Detecting ChatGPT-generated essays in a large-scale writing assessment: Is there a bias against non-native English speakers?” (Computers & Education 217, 2024), drew a sample of essays from the Graduate Record Examinations (GRE) writing assessment and built their own detectors of ChatGPT-generated essays from linguistic features of the ETS e-rater scoring engine and from text-perplexity features. The authors report that these detectors achieved near-perfect detection accuracy and “showed no evidence of bias disadvantaging non-native English speakers.” The study measured detectors the authors constructed for that assessment, not the publicly available tools Liang et al. tested, and its finding is stated for those detectors on that data. doi:10.1016/j.compedu.2024.105070. Abstract and introduction read Oct 14, 2026; the full text is paywalled.
- Seven detectors tested on machine-written, altered, and human-written essays — Perkins et al., 2024
Perkins, Roe, Vu, Postma, Hickerson, McGaughran and Khuat, “Simple techniques to bypass GenAI text detectors: implications for inclusive education” (International Journal of Educational Technology in Higher Education 21:53, published Sep 9, 2024), ran essays through seven detectors: Turnitin's AI detector, GPTZero, ZeroGPT, Copyleaks, Crossplag, the GPT-2 Output Detector and GPTKit. The essays were 15 generated by GPT-4, Bard and Claude 2; 89 versions of those essays altered by six simple text manipulations; and 10 human-written controls by the authors, whom the paper describes as academic staff with English as their primary language and Vietnamese undergraduates writing in English as a second language. On the unaltered machine text the detectors' mean accuracy was 39.5%; on the altered text it fell to 22.14%. On the human-written controls the paper reports that “only 67% of the tests were accurate,” and states a false-accusation rate of 15%, with Copyleaks misidentifying 50% of the human-written samples as AI-written while four of the seven detectors misclassified none. The authors conclude that detector results “should not be used for punitive actions or in accusations against students without a high degree of certainty.” The control set is ten essays, and its figures are stated for that set. Open access (CC BY 4.0). doi:10.1186/s41239-024-00487-w. Retrieved Oct 15, 2026.
- Fourteen tools tested, false-accusation risk computed per tool — Weber-Wulff et al., 2023
“Testing of detection tools for AI-generated text” (International Journal for Educational Integrity, published Dec 25, 2023) tested twelve publicly available tools and two commercial systems, Turnitin and PlagiarismCheck, and concluded the tools were “neither accurate nor reliable” — all scored below 80% accuracy. The study computed, for each tool, the likelihood that a classification would falsely accuse a student: zero for half of the fourteen tools, while six produced false positives, and for one tool half of its positive classifications would have been false accusations. On human-written English text the tools averaged 96% accuracy; on human-written text machine-translated into English, accuracy dropped by 20 percentage points — human writing in another language, translated, read to the tools as more machine-like. doi:10.1007/s40979-023-00146-z. Retrieved Sep 20, 2026.
- The vendor's own bias evaluation — Turnitin, October 2023
A vendor statement about its own product: Turnitin (blog, Oct 26, 2023) reported testing its detector on samples of up to 2,000 texts per group from open datasets of English-as-a-second-language (L2) and first-language (L1) writers, most of them short persuasive or informative writing tasks. For documents of 300 words or more, Turnitin reported that the difference in false-positive rate between L2 and L1 writers was “small, and not statistically significant,” and that each group's rate, while slightly above its 1% target, was not significantly different from it. For documents of 150 to 299 words, the difference between the groups was greater, and the rate was significantly above the target. Turnitin notes that its detector was not among those Liang et al. tested and that it does not score documents under 300 words. Turnitin, Oct 2023. Retrieved Oct 14, 2026.
- Seven detectors tested on non-native English writing — Liang et al., 2023
An Opinion article: “GPT detectors are biased against non-native English writers” (Patterns, Jul 14, 2023) summarizes results the authors first reported in a preprint (arXiv:2304.02819). It reports that they evaluated seven widely used GPT detectors on 91 human-written TOEFL essays from a Chinese forum and 88 US eighth-grade essays from the Hewlett Foundation's ASAP dataset. The detectors classified the US essays accurately but labeled more than half of the TOEFL essays AI-generated — an average false-positive rate of 61.3%, with 19.8% of the essays flagged unanimously by all seven detectors and at least one detector flagging 97.8% of them. When the researchers enriched the TOEFL essays' word choice, the average false-positive rate fell to 11.6%; when they simplified the vocabulary of the US essays, misclassification rose. The authors attribute the pattern to detectors' reliance on text perplexity, which reads a limited range of expression — common in non-native writing — as machine-like. doi:10.1016/j.patter.2023.100779. Retrieved Sep 20, 2026.
- The vendor's own published rates — Turnitin, June 2023
A vendor statement about its own product: Turnitin's Chief Product Officer wrote (blog, Jun 14, 2023) that the document-level false-positive rate is “less than 1% for documents with 20% or more AI writing,” and that the sentence-level false-positive rate — the figure governing any single highlighted passage — is “around 4%,” with such errors more common in documents mixing human and AI writing, at the transitions. Turnitin, 2023. Retrieved Sep 20, 2026.
What the register shows, stated at its scope: false positives are documented in every independent test that measured them; measured rates vary widely with the tool, the dataset, and the writer; and the writers most affected in the research — non-native English writers, English-language learners — are often those least positioned to contest a flag.
What institutions themselves publish about these scores
The institution pages record, campus by campus and in each school's own words, which detection tools a school documents and what weight it says their output carries. A recurring pattern in the collection is an institution stating on its own pages that a detector score alone does not settle a case. The collection page counts, across every institution's detection-status line, how many record such a statement and lists each institution behind the count; among them are Georgia State, Clemson, Northern Arizona, Temple, and LSU. The quotations, their dates, and their sources live on those pages; whether any statement applies to a particular case is for the institution and its process to determine.
Where the records meet the research
Every instrument in the register above reads the submitted text, or its typing rhythm — none of them reads the process by which a piece of writing came to exist. Drafts, version histories, notes, and the people and systems that observed the work happening are records a classifier never sees and cannot score. Preserving and organizing them is covered step by step in the records checklist; capturing the pages a school publishes, as they read today, is covered at what to do with these sources.
How this register stays current
New research is found by citation, not by keyword: a scheduled scan asks the OpenAlex scholarly index for works citing the register's seed studies, which surfaces the literature on detector reliability with far more precision than a search. The scan produces a candidate list and nothing else. A work enters the register only after a person has read it and it meets the register's standing test: it is a primary source — a study, a preprint, or a vendor's statement about its own product; it bears on detector behavior when reading human-written text; and its figures can be read from the work itself. Works the scan surfaces that fail the test are recorded as seen, so a quiet month means no qualifying research was published, not that nothing was checked. Additions and corrections appear in the change log below, on the record.
Revision history of this page
2026-09-20 — Page created. Register entries verified against their originals this date: the Weber-Wulff article read in full at link.springer.com (browser read; the host answers plain fetches with a challenge page), the Liang article at cell.com (browser read), the Turnitin statement at turnitin.com (browser read), and both arXiv abstracts by direct fetch. The Common Sense Media entry is carried at the figures published on this site's homepage; its host did not answer a plain fetch this date and a first-party re-verification is noted for the next review pass. Deliberate omission: the University of San Diego law library guide cited on the homepage as a secondary pointer is not entered here, because the register holds primary sources only. The citation scan (tools/research-scan.py) was seeded this date: 1,126 citing works recorded as seen; triage of that backlog for qualifying entries is pending and will be folded in on the record. Reviewer: Carrie Schluter (review confirmed 2026-09-20).
2026-10-14 — Register reviewed and extended. Every figure in the existing entries was re-checked against its source this date: Liang et al. at PubMed Central, Weber-Wulff et al. at arXiv (v2, the submitted version; the published copy at link.springer.com answered a plain fetch with a challenge page), Stowe et al. (v2, Apr 23, 2026) and the keystroke preprint (v1) at arXiv, and the Turnitin statement at turnitin.com. All figures held. The Common Sense Media entry, carried since 2026-09-20 at the homepage's figures, was verified first-party against the published report (PDF), whose chart gives 20% of Black, 10% of Latino and 7% of White teens; retrieval date added. Wording corrected in the keystroke entry: “more than 99.8%” now reads “99.8% or more,” matching the paper's ≥. Two entries added: Park, Jeong and Kim (arXiv:2608.26710, Aug 27, 2026; 13 detectors, 135,389 document pairs, false-positive rates 0.0% to 100.0%), and Turnitin's own ELL-bias evaluation (blog, Oct 26, 2023). Deferred, not entered: Jiang, Hao, Fauss and Li, “Detecting ChatGPT-generated essays in a large-scale writing assessment: Is there a bias against non-native English speakers?” (Computers & Education 217, 2024; doi:10.1016/j.compedu.2024.105070). The paper bears directly on this register, but it is closed-access and sciencedirect.com refused both a plain fetch and a browser read this date, so its findings are not stated here until the paper is read. Reviewer: Carrie Schluter (review confirmed 2026-10-14).
2026-10-14 — Liang et al. entry labeled as an Opinion article. The Patterns piece is marked “Opinion” on its first page and summarizes results the authors first reported in their arXiv preprint (2304.02819), which it cites as its reference 5; the entry now says so and links the preprint, as the register already labels preprints and vendor statements. No figure changed. Read from the published PDF supplied by the owner this date. Reviewer: Carrie Schluter (review confirmed 2026-10-14).
2026-10-14 — Jiang et al. (2024) entered; the deferral recorded earlier this date is resolved. sciencedirect.com refuses automated reads, so the owner copied the publisher's abstract and the paper's introduction from the article page and supplied them this date. The entry rests on those two sections only: the authors built detectors from e-rater linguistic features and text-perplexity features on GRE writing-assessment essays and report near-perfect accuracy with “no evidence of bias disadvantaging non-native English speakers.” The abstract gives no sample size or error rates, and none is stated here. The entry says the study measured the authors' own detectors, not publicly available tools. Reviewer: Carrie Schluter (review confirmed 2026-10-14).
2026-10-15 — Perkins et al. (2024) entered. The ledger had carried the work as a candidate since the 2026-09-20 seeding; the owner raised it this date from a secondary summary, which is not entered. The paper itself was read at the publisher's open-access PDF (educationaltechnologyjournal.springeropen.com) this date and every figure in the entry is taken from it: seven detectors; 15 machine-written, 89 altered and 10 human-written samples; mean accuracy 39.5% on unaltered machine text, 22.14% on altered text; “only 67% of the tests were accurate” on the human-written controls, a false-accusation rate the paper states as 15%, Copyleaks at 50% of human samples misidentified, four of seven detectors with none. The six alteration techniques are not named here, per this page's rule that an evasion finding is reported as a reliability finding and never as a method. One internal inconsistency in the paper is noted for the reviewer and not resolved on the page: it describes the tools as “six detection tools” in the testing protocol while listing and tabulating seven, and it does not state the denominator behind the 15% figure. Section 02 now points to the collection page's published-cautions count (/collection/#caution) as the place where the pattern it describes is counted and listed; the five named examples remain. Reviewer: Carrie Schluter (review confirmed 2026-10-15).
Spot an error or an outdated quotation? hello@gatheredwork.com — corrections are made on the record, in this log.