The Evolution of Assessment in STEM Education Amidst the Generative AI Era: A Systematic Review of Pedagogical Shifts and Academic Integrity

The proliferation of Generative Artificial Intelligence (GenAI), including large language models (LLMs) such as ChatGPT, has fundamentally destabilized traditional assessment practices in STEM education by rendering correctness-based evaluation instruments increasingly invalid. This systematic literature review, conducted following the PRISMA 2020 protocol, synthesizes 52 peer-reviewed articles published between 2022 and 2026 to map assessment responses across STEM disciplines, with particular concentration in computing, engineering, physics, and mathematics, in response to GenAI adoption. A comprehensive Boolean search was executed across Scopus, Web of Science, and ERIC, and a seven-criterion quality appraisal instrument adapted from the Mixed Methods Appraisal Tool (MMAT) was applied to ensure methodological rigour. Bibliometric analysis using VOSviewer reveals that large language models constitute the dominant conceptual node of the field, co-occurring with academic integrity and engineering education as the primary disciplinary contexts. Thematic synthesis yielded three pedagogical response paradigms, organized into the ARIA Framework: AI-Resistance, emphasizing oral examinations and process portfolios; AI-Incorporated, embedding GenAI as a pedagogical scaffold while evaluating students' critical engagement; and AI-Leading, positioning students as auditors of AI-generated artefacts. AI-detection software was consistently critiqued as unreliable and inequitable, particularly for non-native English speakers. Critical thinking was reconceptualized as comprising three relational competencies: AI Output Literacy, Epistemic Autonomy, and Collaborative Orchestration. This review contributes the ARIA Framework as an analytical classification tool and identifies priority research gaps including longitudinal efficacy, K-12 contexts, disciplinary breadth beyond computing and engineering, and cross-cultural applicability.