Abstract
<title>Abstract</title> <p>Artificial intelligence in STEM assessment routinely prioritizes computational efficiency over equity. This oversight silently codes historical demographic biases into modern evaluation tools. The objective of this study is to reengineer automated assessment, shifting its operational mandate from uncritical algorithmic optimization to active bias mitigation. Methodologically, the research employs an integrative literature review and thematic analysis. The study evaluates a purposive sample of fifty-seven academic papers to map the exact intersection of technical capability and educational justice. The primary result of this analysis is the development of the AI-Driven Assessment in STEM Education Conceptual Framework for Educational Equity (ASEAF). The findings reveal a stark reality. Computational accuracy alone cannot achieve educational fairness without structural intervention. To resolve this, the ASEAF framework is designed as a cyclical sociotechnical ecosystem built on three interdependent pillars. First, a Technical-Pedagogical Pipeline generates adaptive cognitive profiling. Second, a Bias-Auditing Interface intercepts this data to enforce error rate parity and demographic calibration. Third, a Governance-Oversight Mechanism mandates human-in-the-loop validation, ensuring algorithms remain strictly accountable to educators. Ultimately, this framework provides a practical blueprint for higher education institutions. It requires them to transition from passive consumers of proprietary software into active algorithmic auditors, ensuring predictive analytics dismantle systemic inequities rather than quietly reinforcing them.</p>