Modeling university teachers’ AI adoption and practice trajectories using longitudinal learning analytics: a three-semester observational study

The higher education sector requires empirical evidence on how artificial intelligence (AI) can use longitudinal learning analytics on digital trace data. This study developed and tested a longitudinal learning analytics strategy to model teacher AI adoption and trace-based teaching-practice trajectories over three semesters, identify actionable predictors, and inform scalable, equity-sensitive monitoring and support. Data were collected from 312 faculty members across three Chinese universities through digital traces of learning-management logs, automated assessments, and tutorials. Instructors in the higher-engagement cohort increased their routine use of intelligent tools, diversified workflows, and reduced adaptation time, which was associated with improved course indicators. Three signal events per week, portfolio diversity, and adaptation latency were the strongest statistical predictors of outcome change. Teaching activity rose by 4.4 events per week, pass rates increased by 3.5 percentage points, and a large time-by-cohort interaction was observed (F(1,308) = 96.8; η² = 0.239). Heat maps revealed a lasting shift from passive to active profiles, while radar profiles underscored the importance of intensity, diversity, and adaptation speed. Institutions can apply these findings by monitoring the three signals to enhance faculty practice. Practice indicators is conceptualized here as an emergent, trace-observed change in teaching practice rather than a standardized intervention or dosage-delivered program.