The emergence of large language models has precipitated a fundamental disruption to higher education assessment. Conventional instruments are susceptible to AI-assisted completion, threatening construct validity across disciplines and institutional contexts. When submitted work reflects AI capability rather than student competency, the inferential chain from performance to qualification is invalidated – a problem of design rather than detection. This systematic review makes two contributions. First, it theorises and grounds the distinction between AI-resistant assessment, which seeks to prevent or impede AI access through containment, and AI-resilient assessment, which assesses genuine human cognitive performance of intrinsic educational value regardless of the AI tools that exist. Second, it synthesises peer-reviewed empirical literature, policy documentation, and sector guidance published between 2022 and 2025. A PRISMA 2020-compliant search of five databases was conducted; two independent reviewers screened all records with substantial inter-rater agreement. Forty-seven peer-reviewed studies, 23 policy documents, and 11 sector reports were included. Six evidence-based assessment categories were identified, taxonomised, and evaluated against a structured adversarial threat taxonomy. AI-detection tools are structurally insufficient as primary responses. A Six-Layer AI-Resilient Assessment Stack, grounded in multi-trait multi-method validation logic, is proposed as an integrated institutional framework. Persistent challenges include equity, reliability, faculty development, and data protection.
Asrat Genet Amnie (Fri,) studied this question.