AI-Powered Assessment: A Shift from Closed-Book Examinations to Skill-Based Learning

Traditional closed-book examinations assess knowledge recall under artificial information-isolation constraints that are structurally misaligned with professional competency requirements, produce high-stakes anxiety that suppresses authentic performance, and generate assessment data insufficient to inform personalised learning interventions at scale. This paper proposes an AI-Powered Skill-Based Assessment Framework (APSAF) integrating an automated essay and multi-modal response analysis engine, an adaptive skill assessment and learning analytics module, and an anti-plagiarism, assessment integrity, and DEI compliance gateway, evaluated across 56 educational institutions (universities$\mathbf{n} \boldsymbol{=} \mathbf{2 2}$, secondary schools$\mathbf{n} \boldsymbol{=} \mathbf{1 8}$, professional certification bodies$\mathbf{n} \boldsymbol{=} \mathbf{1 6} \boldsymbol{)}$processing 3,280,000 assessment records (2,296,000 training / 984,000 test, 70/30 stratified split) and 284,200 assessment events across 6 subject domains over a 38-month evaluation horizon. The APSAF achieves a composite assessment effectiveness score of r = 0.946 against expert-validated learning outcome benchmarks, outperforming MLonly assessment models$(\mathbf{r}=\mathbf{0. 9 0 0}$; +5.1%$[(0.946-0.900) / 0.900])$, rule-based assessment systems$(\mathbf{r}=\mathbf{0. 5 7 8}$; +63.7% [(0.946-0.578)/0.578]), and traditional closed-book examination scoring$(\mathbf{r}=\mathbf{0. 4 9 4}; +\mathbf{9 1. 5 \%}[(\mathbf{0. 9 4 6 - 0. 4 9 4}) / \mathbf{0. 4 9 4}])$. Deployment outcomes: +64.8% assessment accuracy, +46.8% student learning outcome prediction accuracy, −56.2% educator marking time, +37.2% personalised feedback quality, 93.6% assessment integrity rate, and +52.8% skill-based competency alignment. Sub-domain performance: Automated Essay Scoring (r = 0.968), Adaptive Skill Assessment (0.964), Real-Time Feedback and Learning Analytics (0.952), and Anti-Plagiarism and Assessment Integrity (0.932). Cultural and linguistic bias in automated scoring degrades APSAF accuracy to 0.933 (-1.4% [(0.946-0.933)/0.946]), and sparse training data in niche subject domains to 0.919 (-2.9% [(0.946-0.919)/0.946]), with mitigation protocols for both conditions.

Hali tarjima qilinmagan