Beyond the detector trap: a human-centred evidentiary framework for fair academic integrity decisions in the generative AI era

Abstract

Generative AI (GenAI) and large language models have disrupted traditional assumptions about academic writing, authorship, and assessment in higher education. While institutions often rely on AI-detection tools, this approach risks a detector trap: reducing complex academic integrity judgements to mere algorithmic classification of textual origin. This article argues that the central problem is not simply whether AI is present in a submission, but whether verifiable, procedurally fair evidence demonstrates misconduct or a breach of academic standards. Detector scores may prompt closer review, but they must not independently justify disciplinary sanctions. To address this challenge, the article proposes a human-centred evidentiary framework for investigating suspected GenAI misuse. This framework progresses from initial signals to final decisions through four distinct analytic layers: text-internal anomalies, algorithmic indicators, epistemic verification, and structured student-educator dialogue. It classifies concerns by evidentiary reliability into low, medium, or high categories, matching each category with proportionate pedagogical, procedural, or formal institutional responses. By shifting the focus from detection accuracy to defensible decision-making, this article contributes to scholarship and practice in higher education assessment and academic integrity. It concludes that fair institutional practice requires explicit AI-use policies, transparent evidentiary standards, structured opportunities for student response, and innovative assessment designs.