An AI-writing detector should not decide whether a student cheated
A peer-reviewed review of 90 studies found that false-positive rates vary too much across tools, text types and testing conditions to support a single reliable figure. The practical safeguard is procedural: detector scores need contextual evidence, human review and an appeal route.
The short version
- What happened
- The systematic review included 90 studies published from 2020 to 2026: 81 indexed in Scopus and nine found through Google Scholar.
- Who it affects
- Students can face serious academic and psychological consequences when authentic work is misclassified; multilingual writers and people using concise, formulaic academic styles may be especially exposed.
- What remains unproven
- The review covered English-language publications found mainly through Scopus plus Google Scholar, did not pool false-positive rates, and combined studies of different detector versions, thresholds, datasets and educational settings.
Keep in mind
The review covered English-language publications found mainly through Scopus plus Google Scholar, did not pool false-positive rates, and combined studies of different detector versions, thresholds, datasets and educational settings.
Source analysis · 1 linked source. Primary analysed source date: 17 October 2026.
Read the full report · 4 min The evidence, methods, implications and limitations
A score is a prompt to investigate, not a verdict
Universities should not use an AI-writing detector score as definitive evidence that a student cheated, according to a new peer-reviewed systematic review. Across 90 studies, false-positive results varied substantially between products, model versions, thresholds, datasets, languages and assignment types. That variation prevented the authors from calculating a single pooled false-positive rate. The clearest conclusion is therefore procedural: an automated flag cannot carry the same weight as verified authorship evidence.[1]
Who is most exposed when the tool is wrong
A mistaken flag can trigger an investigation, a grade penalty, stress and reputational harm. The review found recurring concerns for multilingual and non-native-English writers, whose legitimate language patterns may be under-represented in detector training data. Short, formal or highly structured writing can also resemble the surface patterns detectors associate with generated text. For teachers and academic-integrity teams, the practical response is to examine drafts, references, version history and the student's explanation, publish the limits of any tool being used, and provide a route to challenge the result.[1]
The distinction that matters
How a detector flag should move through a fair process
The review supports a human-in-the-loop pathway instead of an automatic misconduct decision.
Flag
A detector marks text as possibly generated
Treat the score as uncertain information
Review
Check drafts, sources, context and the student's account
Look for corroborating evidence
Decision
Use a transparent policy and appeal route
Do not punish on the score alone
What the review did—and what remains unproven
The authors searched Scopus using terms for AI-text detection and false classifications, retrieving 120 records before date, subject and relevance screening left 81. A supplementary Google Scholar search found 18 candidates and retained nine, producing a final set of 90 English-language studies from 2020 to 2026. They appraised reporting quality, used VOSviewer for keyword mapping, structural topic modelling for themes and qualitative synthesis for causes, consequences and safeguards. The analysis drew partly on abstracts and author keywords rather than every study's full numerical result.
This is a synthesis of heterogeneous research, not an independent head-to-head trial of current commercial detectors. Earlier product versions may no longer behave the same way, publication bias is possible, and studies outside English or the two searched sources may have been missed. The proposed safeguards—human review, AI literacy, transparent policy and appeals—are evidence-informed, but the review says their effectiveness still needs real-world testing. Confidence would rise with preregistered evaluations using current tool versions, representative student writing and measured outcomes from actual institutional decision processes.[1]