Topic
A score from an AI-writing detector is an estimate about a piece of text. It is not a record of how the text was made, and it says nothing about intent. Anyone weighing such a score next to a student's work should first know what published tests of these tools report.
Some papers here evaluate tools, each on its own set of texts. A test of many tools judged them inaccurate and unreliable on its document set. A comparison of detectors reports results that differed a lot between tools and chatbot versions. A 2026 evaluation of two commercial detectors raises a possible fairness concern for English-as-a-foreign-language writing, which it calls a borderline trend. On studies like these, a detector's output is at most a reason to look further.
Every one of these results belongs to particular tool versions and test texts, which may differ from the coursework in front of a marker.
International Journal for Educational IntegrityTool evaluation
Can current tools tell human writing from chatbot text? On their own document set, the authors judge the tools they tried inaccurate and unreliable, inclined to label text human, and weaker after obfuscation.
For teaching A caution against treating a tool's output as proof about any one student, based on the authors' own document set.
Core reading Free to read
International Journal for Educational IntegrityTool evaluation
Detector results on chatbot paragraphs about an engineering topic were better for an older model than a newer one, the authors report, and inconsistent on human-written controls, with some false positives.
For teaching From a small single-topic test, a reason for caution before reading much into a detector score on any one student's work.
Core reading Free to read
International Journal for Educational IntegrityTool evaluation
On the authors' set of 192 texts, including pre-AI writing by English-as-a-foreign-language students, two commercial detectors reached overall accuracy of 0.69 and 0.61, did poorly on mixed texts, and showed a borderline fairness signal.
For teaching Suggests a detector's output is at most a reason to look further, and the authors raise a possible fairness concern for second-language student writing.
Core reading Free to read
Journal of Academic EthicsExperiment
The authors report that an AI-writing detector flagged 91% of their 22 ChatGPT-written test submissions, but only 54.8% of the AI text in them; the 15 staff marking them referred 54.5% for misconduct.
For teaching May inform thinking on assessment design and on how staff read detector reports, from one experiment with 22 submissions and 15 markers.
Core reading Free to read
Journal of University Teaching and Learning PracticeCommentary
Where does student use of AI writing tools become a breach of integrity? The paper argues the deciding factor is whether students make their use clear, with institutional policies setting what counts as misconduct.
For teaching Points towards clear disclosure expectations in course policies, as an argument rather than tested evidence.
Free to read
Open Information ScienceTool evaluation
On its own set of first-year composition essays and chatbot essays, three of 16 AI text detectors scored well throughout, while most others struggled to separate student essays from those of a newer chatbot.
For teaching Results differed a lot between tools and chatbot versions, so a flag from any one tool needs care; this rests on a single test set.
Free to read
Journal of Applied Learning & TeachingReview
Pulling together 17 articles from 2023 on how well AI-writing detectors told machine text from human text, the review reports the studies found every tool inconsistent and recommends adding human reviewers.
For teaching Human review alongside any detector result is the review's recommendation, since the reviewed studies found every tool they tested inconsistent.
Free to read
PLoS ONESurvey
Surveyed academics differed on when AI help with a manuscript should be reported, and one detector often flagged abstracts given only grammar fixes. The authors suggest reporting rules may be hard to enforce with detectors.
For teaching Offers a reminder that tool flags may not separate light editing from substantial AI writing, though the setting here is journal manuscripts rather than coursework.
Free to read
American Journal of IR 4 0 and BeyondReview
Across 34 reviewed articles, the authors report that most AI-text detectors scored above 50% accuracy in the studies yet were judged unreliable, with fairness concerns for writers whose first language is not English.
For teaching The review urges human judgement alongside any detector and notes fairness concerns for writers whose first language is not English, so a detector result needs careful weighing.
Free to read
European Journal of Engineering EducationCase study
Telling machine-generated article writing from human writing was hard for academics new to such generators, the authors report from a case study, and a survey suggested that awareness training improved their judgement.
For teaching Awareness sessions on article generators may help staff judge suspect writing, going by one case study and survey.
Free to read
Accountability in ResearchTool evaluation
How accurate are AI text detectors on behavioural health writing? The authors report problematic error rates for free and paid tools, with a free one marking a median 27.2% of academic text as AI-generated.
For teaching Detector scores on scholarly writing may call for cautious handling, on evidence limited to published behavioural health articles.