A ChatGPT-4o essay-detection test highlights false-positive risks

Updated Sep 07, 2026

A model that recognises some AI-assisted writing can still falsely accuse human writers. Qinghao Guan and Yangxi Han report this problem when testing ChatGPT-4o as an essay detector.

What the study reports

The study involved 156 STEM students at one university in central China, in a second-language writing context. It combined a survey of AI-related views with a test set of 78 human-written and 78 ChatGPT-assisted essays.

The publisher’s results section reports that ChatGPT-4o identified 63 of the 78 assisted essays as AI-assisted. It also incorrectly labelled 53 of the 78 human-written essays as AI-generated. These are results for this model, task and dataset, not a benchmark for every AI-detection product or later model version.

The authors discuss ethical concerns, including privacy and the risk of misjudging students’ work. Accessible methods and results support this account; complete survey tables and the classification figure were not retrieved for this review. Earlier detailed attitude percentages and an unverified quotation have been removed.

Questions for assessment policy

As editorial recommendations, do not treat a detector flag as proof of authorship or misconduct. Any use needs an appropriate evidence process, clear communication about permitted assistance and a fair opportunity for a student to respond.

Evaluate the actual tool and setting under consideration. Ask students about their experience of the process, but assess detection performance independently of whether a tool is popular or easy to use. Protect submitted work according to the institution’s agreed data arrangements.

Reference

Qinghao Guan and Yangxi Han. “From AI to authorship: Exploring the use of LLM detection tools for calling on ‘originality’ of students in academic environments.” Innovations in Education and Teaching International. DOI: 10.1080/14703297.2025.2511062.

Reviewed 7 September 2026 against accessible publisher sections. The headline now identifies the tested model, and unverified survey details and quotation have been removed. This is a limited-access review.