Experience Beats Equipment: What 1,000 Indonesian Teachers Reveal About AI in Inclusive Classrooms

Artificial intelligence has been heralded as the great equalizer of modern education, a technology capable of tailoring lessons to every learner regardless of ability. But a sweeping new study from Indonesia suggests that the decisive factor in whether AI truly transforms classrooms is not the sophistication of the software or the availability of devices. It is the lived, hands-on experience of the teachers standing at the front of the room. In one of the largest investigations of its kind in a developing country, researchers surveyed 1,000 teachers in inclusive primary and secondary schools across the Indonesian archipelago and found that professional experience with AI exerted a stronger influence on teachers’ attitudes than the quality of AI integration itself.

The research, published in Discover Education, was led by I Wayan Sumandya of Universitas PGRI Mahadewa Indonesia together with colleagues from Universitas Pendidikan Ganesha and other institutions. The team set out to answer a deceptively simple question: what actually shapes whether teachers in inclusive schools, where students with diverse learning needs and disabilities learn alongside their peers, come to see AI as a valuable pedagogical ally? To find out, they built a predictive statistical model that tested two candidate forces simultaneously: the quality with which AI was woven into STEM teaching, and the depth of teachers’ own professional experience with the technology.

The context matters enormously. Indonesia is a maritime nation of more than 17,000 islands, where disparities in internet access, digital infrastructure, and teacher training are stark between urban centers and remote regions. Inclusive schools there face a multidimensional challenge: teachers must integrate emerging technologies while simultaneously ensuring accessibility, fairness, and differentiated instruction for students with visual, hearing, cognitive, and other impairments. AI tools promise real benefits in this setting. Speech-to-text and text-to-speech systems can assist students with sensory impairments, adaptive platforms can adjust content in real time to individual learning paces, and predictive analytics can flag students at risk of falling behind before it is too late to intervene.

Yet most previous research on AI in education has been what the authors call technocentric, focusing on algorithmic capabilities, student-facing applications, and theoretical potential while treating teachers as passive users rather than active agents of change. Established frameworks such as the Technology Acceptance Model, which holds that perceived usefulness and ease of use drive adoption, and the Technological Pedagogical Content Knowledge framework have tended to treat teacher-related variables as static background factors rather than dynamic constructs shaped by direct experience. The Indonesian team argued that for inclusive classrooms, teachers’ perceptions are forged in the crucible of lived experience: successes, failures, technical glitches, and moments when AI genuinely helped a struggling student grasp a difficult concept.

To test this idea rigorously, the researchers employed a quantitative explanatory survey design with purposive sampling. They recruited active teachers with at least two years of experience, participation in AI or STEM-related training, and direct involvement in inclusive teaching. Questionnaires were distributed through provincial teacher associations, STEM education communities, and inclusive education networks across Bali, Java, Sumatra, Sulawesi, and Nusa Tenggara. Of 1,183 initial responses, 1,000 complete surveys survived data screening, a response rate of 84.53 percent. Respondents answered items on five-point Likert scales covering three constructs: AI integration in STEM teaching, encompassing planning, tool use, student engagement, and data-driven assessment; teachers’ experience, including technological competency, challenges encountered, and institutional support; and teachers’ perspectives, covering perceived benefits, implementation readiness, ethical considerations, and future expectations.

The sample was demographically broad. Women made up 62.4 percent of respondents, and teachers were distributed across primary schools (41.8 percent), junior secondary (34.6 percent), and senior secondary levels (23.6 percent). Nearly half had six to fifteen years of teaching experience, a quarter had more than fifteen years, and 71.2 percent had previously participated in AI or STEM professional development. The instrument was refined through expert review by three specialists and piloted with 30 teachers, achieving Cronbach’s Alpha reliability values between 0.82 and 0.89 before the main survey launched.

Analysis was conducted using Partial Least Squares Structural Equation Modeling, or PLS-SEM, a variance-based technique well suited to predictive models in emerging research areas. The measurement models proved robust: all indicator loadings exceeded 0.70, composite reliability values ranged from 0.906 to 0.918, and average variance extracted surpassed 0.65 for every construct. Discriminant validity was confirmed through both the Fornell-Larcker criteria and the HTMT ratio, all comfortably below the 0.90 threshold, establishing that the three constructs, while related, were statistically distinct. Multicollinearity checks showed inner VIF values below 5, and common method bias was ruled out through full collinearity VIF values under 3.3 alongside procedural safeguards such as anonymous participation.

The structural results were striking. Both hypotheses were supported, but with a telling asymmetry. AI integration quality significantly influenced teachers’ perceptions, with a path coefficient of 0.399 and a t-statistic of 17.424, well beyond the significance threshold. Yet teachers’ professional experience proved the stronger predictor, with a coefficient of 0.512 and a t-statistic of 24.121. Effect sizes told the same story: experience produced a substantial effect (f-squared of 0.36, above the 0.35 benchmark for a large effect), while integration quality showed a moderate effect of 0.21. Together the two factors explained 44.9 percent of the variance in teachers’ perceptions, a moderate level of explanatory power, and the model achieved satisfactory fit with an SRMR of 0.045, well under the 0.08 criterion.

Perhaps the most innovative contribution lies in the model’s demonstrated predictive power. Using PLS Predict analysis, the team compared their model’s out-of-sample error against a naive linear regression benchmark. The PLS-SEM model posted lower root mean squared error and mean absolute error values across all five indicators of teachers’ perspectives, and all Q-squared predicted values were positive, confirming genuine predictive relevance. In practical terms, this means the model does not merely describe the present; it can simulate future scenarios. Policymakers could, for instance, estimate how expanding experiential training programs might shift teachers’ implementation readiness or long-term expectations, turning the study from a snapshot into a planning instrument.

The theoretical and practical implications ripple outward from Indonesia to any education system attempting to balance digital transformation with equity. The findings extend the Technology Acceptance Model by positioning teachers’ experience as a critical antecedent, stronger than system integration alone, in complex inclusive settings. They align with constructivist learning theory, which holds that knowledge and attitudes are built through experience and reflection, and with prior work showing that hands-on engagement reduces technology anxiety and builds self-efficacy. The authors recommend that schools move beyond one-way theoretical training toward immersive, ongoing professional development in which teachers can experiment with AI, make mistakes, and reflect in a supportive environment. Institutional leaders should cultivate professional learning communities where educators share implementation strategies and accessibility adaptations, while policymakers should fund sustainable mentoring, accessibility support systems, and inclusive AI governance frameworks rather than hardware alone. The study’s limitations, including purposive sampling, a cross-sectional design, and self-reported data, mean the findings are associative rather than definitively causal, and the authors call for longitudinal and mixed-methods research to track how experience compounds over time. But the core message is already clear: in the race to bring AI into classrooms that serve every learner, the technology is only half the equation. The other half is the teacher who has actually used it, struggled with it, and seen it work.

Subject of Research: Teacher experience and AI integration in inclusive STEM education in Indonesia

Article Title: AI integration in STEM learning in inclusive schools in Indonesia from teachers experiences and perspectives

Article References: AI integration in STEM learning in inclusive schools in Indonesia from teachers experiences and perspectives. (n.d.). https://doi.org/10.1007/s44217-026-02139-9

Image Credits: AI Generated

DOI: 10.1007/s44217-026-02139-9

Keywords: artificial intelligence, inclusive education, STEM education, teacher perception, PLS-SEM, Indonesia, educational technology, teacher professional development, technology acceptance model, accessibility, adaptive learning, survey research