What counts as evidence of learning when students use generative AI? A critical integrative review of assessment validity in higher education

Abstract

Generative artificial intelligence (GenAI) can support feedback, practice, and drafting, but it weakens a familiar assumption of take-home assessment: that the submitted product reliably records students' own learning. Higher education now offers many AI-aware assessment proposals, yet their recommendations often overlap, and the evidential purpose of each safeguard remains unclear. This critical integrative review examined 31 verified works published from 2021 to August 2026, comprising 26 primary empirical, technical, implementation, or document-analysis studies and five review or conceptual papers. Sources were coded for the learning claim at stake, the role allowed for GenAI, the evidence collected, the inference supported, the implementation burden, and the equity risk. The review found that detector scores cannot establish learning, authentic tasks alone remain vulnerable, and no single process record proves independent competence. Stronger designs combine evidence channels: a final product, selected process traces, a short oral or live check, and, where retention matters, delayed unaided performance. Building on this finding, the article presents an evidence-calibration model that asks instructors to define the learning claim, locate the AI-sensitive link in the inference, choose the smallest useful evidence bundle, and match verification intensity to stakes, workload, access, and student rights. The model is a design aid rather than a validated instrument. It offers a practical basis for assessment decisions while clarifying where further empirical testing is needed. The review contributes by synthesizing overlapping recommendations, making each safeguard's evidential purpose explicit, and foregrounding equity and student rights in AI-aware assessment design