A review of 90 studies links false alarms to language bias, limited context and narrow training data.
Researchers reviewed 90 studies published between 2020 and 2026, including 81 publications indexed in Scopus and nine identified through Google Scholar. They combined bibliometric mapping and topic modelling with a qualitative review of the causes, effects and possible responses to false-positive AI detection in education.
The studies covered academic integrity, responsible use of AI, detection accuracy and assessment. False-positive rates varied substantially between tools and testing settings, making direct numerical comparisons difficult. The review points to human review, transparent detection practices, AI literacy, clear policies and accessible appeals as ways to reduce harm from incorrect classifications.
What the review found
The review identified three broad areas of research: academic integrity and student assessment; plagiarism and responsible AI governance; and detection accuracy, reliability and false positives.
Across the literature, false-positive outcomes differed widely depending on the tool and evaluation setting. The review identified several factors associated with incorrect flags of human-written work: algorithmic and language bias, limited understanding of context, reliance on surface-level language features and a lack of diversity in training data. The reported consequences included risks to assessment fairness, student confidence and institutional decision-making.
Evidence and caveats
This was a systematic review of 90 eligible studies published from 2020 through 2026, supported by bibliometric mapping, structural topic modelling and qualitative synthesis. It brings together evidence about causes, consequences and possible safeguards rather than testing one detector in a single controlled evaluation.
The review reports that false-positive results vary substantially across tools and evaluation settings, so the findings do not provide one general false-positive rate for all detectors. The abstract also does not establish that every detector is equally affected or show how well the proposed mitigation strategies work in practice.
