Training the Trainer: Measuring the Long-Term Effectiveness of AI Upskilling Programs for Indonesian Educators

Abstract

Indonesia's national artificial intelligence (AI) education roadmap (2025–2045) positions teacher AI competency as foundational to its Golden Generation vision, yet the long-term effectiveness of AI upskilling programs for Indonesian educators remains empirically unexamined. This study aims to evaluate the extent to which AI-related competencies are retained over time, identify key factors influencing sustained AI integration among teachers, and compare the effectiveness of different upskilling programs across regional contexts. This longitudinal mixed-methods study measured the durability of AI competency gains among 347 in-service K–12 teachers across urban, peri-urban, and remote regional strata over four measurement waves: pre-intervention, immediate post-intervention, six-month follow-up, and twelve-month follow-up. Participants were drawn from two nationally deployed AI upskilling programs  Platform Merdeka Mengajar (PMM) and Microsoft Elevate   and assessed using a validated four-subscale questionnaire, semi-structured interviews, and structured classroom observation checklists. Results revealed a consistent "spike-and-decay" trajectory across all groups, with teachers retaining an average of only 63.5% of post-intervention competency gains at the twelve-month wave. Microsoft Elevate participants demonstrated significantly superior retention compared to PMM participants, and remote-stratum teachers experienced the steepest decay. Post-intervention AI self-efficacy, peer collaboration frequency, and perceived program relevance emerged as the strongest predictors of sustained behavioral integration. Qualitative findings identified infrastructure deficits, institutional isolation, curricular misalignment, and confidence erosion as primary disinvestment mechanisms. The study concludes that sustainable AI teacher competency requires a systems-level response encompassing mandatory post-training support, differentiated regional resourcing, and curriculum-embedded AI learning communities.