AI content detection in student text assignments, false positive rates, and academic integrity

Synthesis: Maghami, Ho, and Camilleri (2026) synthesize 90 studies published between 2020 and 2026 on AI Detection in higher education, 81 retrieved from Scopus and nine added from Google Scholar. They combine VOSviewer keyword mapping with Structural Topic Modeling and a qualitative synthesis, and the corpus resolves into three themes: academic integrity and student assessment, plagiarism and responsible AI governance, and detection accuracy, reliability, and false positives. Their central claim is that reported false-positive rates cannot be pooled, because studies differ in detector, version, threshold, dataset, text length, language, and design, yet the causes recur across them: language and algorithmic bias, limited contextual understanding, reliance on surface-level linguistic features, and narrow training data. The review frames false positives as a AI Governance problem rather than only a technical one, and recommends transparent detection practices, Human-in-the-Loop review, AI Literacy, clear institutional policy, and accessible appeal routes.

Key Findings

1. Ninety studies drawn from two databases. The corpus comprises 81 Scopus-indexed publications and nine studies located through Google Scholar, restricted to English-language work published between 2020 and 2026.
2. Reported false-positive rates resist pooling. Detector type and version, classification threshold, dataset, text length, language, academic context, and experimental design all vary, so the review treats individual rates as context-bound rather than comparable estimates.
3. Four causes recur across the literature. Algorithmic and language bias, limited contextual understanding, dependence on surface-level linguistic indicators such as perplexity, and insufficiently diverse or outdated training data explain most misclassification.
4. Legitimate writing is systematically exposed. Abstracts, literature reviews, examination responses, translated text, and heavily edited or paraphrased work are flagged disproportionately because their concise, uniform, formulaic surface resembles machine output.
5. Small error rates scale into large harms. A 1% false-positive rate applied to an estimated 22.35 million essays submitted annually by first-year United States college students would wrongly flag more than 223,000 essays.
6. Vendor accuracy claims obscure error structure. Providers report accuracy above 95% for Copyleaks, Turnitin, and GPTZero, yet one evaluation of 30 free detectors found only two that classified all student essays correctly.
7. Layered responses outperform better detectors. Aggregating multiple detectors, individually 84% to 98% accurate alongside human raters, reduced false positives to almost zero in one anatomy and physiology study.

What the Review Measured, and Against Which Tools

The evidence base is a PRISMA-guided Meta-Analysis and Systematic Review rather than a meta-analysis. The Scopus query returned 120 records, 116 after the date limit, and 81 after subject-area and relevance screening; the Google Scholar supplement began with 18 records and retained nine. Analysis combined keyword co-occurrence in VOSviewer 1.6.21, with a minimum occurrence threshold of two, and Structural Topic Modeling run first as six topics and then consolidated into three. The tool vintage matters for reading any number in the paper: the studies evaluated Turnitin, GPTZero, Copyleaks, ZeroGPT, GPTRadar, Writer, Originality.AI, the OpenAI classifier, and assorted support vector machine and BERT-based models, applied to output from ChatGPT, GPT-4, Claude, Gemini, and Adobe Firefly. Because the window runs from 2020 to 2026, much of the measured performance belongs to detector generations that have since been retrained, even as the generative models they were built to catch have also changed.

Why Risk Falls Unevenly

The review's equity argument is structural rather than incidental. Detectors estimate authorship from statistical regularity, so writing that is deliberately consistent reads as suspicious. That penalises an abstract's compressed formality, a methods section's repetition, and a nonnative English writer's careful, taught phrasing, which the authors describe as simpler phrasing and structure resembling what tools associate with AI. Translation adds smoothing; grammar and spelling tools remove the very errors detectors treat as human; multiple rounds of legitimate editing flatten tone. Reviewers compound the problem when they read unusual vocabulary or grammatical inconsistency as confirmation of a flag. Because identical thresholds therefore bind different students differently, the review treats Bias Mitigation and Differential Effects Across Learner Groups as questions of Assessment Validity and procedural fairness, not only of model tuning. In the authors' terms, false positives are foreseeable outcomes of deploying detection technology without adequate understanding, oversight, or procedural safeguards.

From Score to Sanction

The paper's most transferable contribution is procedural. A detector score becomes harmful only when an institution converts it into an allegation, and the review catalogues the policy settings that do so: thresholds tuned to catch AI rather than to avoid error, one-size-fits-all rules across disciplines and languages, mass screening in large classes where a small error rate yields many flags, staff treating a high score as proof, absent context checks, and weak transparency or appeal routes. Vendor calibration claims do not neutralise this, since a Bloomberg evaluation of GPTZero and Copyleaks reported false-positive rates of approximately 1% to 2% on essays written before generative AI was widespread. The proposed remedy is a three-part framework in which technological improvement, educational and policy measures, and institutional support systems operate together, with explainable AI positioned as the link between them. Students then argue about evidence rather than about a number.

What this means for practice

• Instructors. Never treat a score as a finding. Require corroboration from drafts, version history, notes, or an oral check-in before raising a concern, and do not run detection on short texts or on the work of multilingual writers, where the review reports higher false-positive risk.
• Administrators. Write the appeal route before you license the tool: named reviewer, documented evidence standard, stated timeline, and a record of the decision, so that a contested flag has a procedure rather than a committee.
• Policymakers. Set the evidentiary standard for misconduct cases centrally and require periodic audits of error rates across student populations, because detector output alone will not meet a standard of proof the review considers defensible.
• Program designers. Move assessment weight toward reasoning, process evidence, and reflection, which the review finds less vulnerable than tasks that mainly require producing text, and teach AI Use and Disclosure Statements alongside AI Literacy so that permitted assistance is visible rather than inferred.

Limitations

• The search covered English-language publications only and relied primarily on Scopus with a Google Scholar supplement, so relevant work in other languages or databases may be missing.
• The synthesis rests on 90 studies whose detector tools, versions, thresholds, datasets, and text types differ substantially, which is why the review reports no pooled false-positive estimate.
• Parts of the bibliometric and thematic analysis drew on abstracts and author keywords rather than full texts, so methodological detail in individual studies may be under-represented, and possible publication bias is acknowledged.
• The review's own Limitations in AIEd Research is temporal: generative AI and detection both moved quickly across 2020 to 2026, and findings for earlier tool versions may not describe current performance.
• Consequence claims come from studies that document accusations and student reports, not from longitudinal tracking, so the review calls for longitudinal work in Research Methods in AIED linking false-positive exposure to confidence, motivation, stress, and institutional Trust.
• No student-level outcome data were analyzed by the review itself, no datasets were generated, and the 2026 topic trends reflect a partial publication year.

Citation

Maghami, M. R., Ho, C. K., & Camilleri, P. (2026). AI content detection in student text assignments, false positive rates, and academic integrity. Discover Education, 5, 1192.