This study examined attitudes toward artificial intelligence (AI) among pre-service primary school teachers in Mongolia and evaluated the psychometric properties of a Mongolian adaptation of the Students’ Attitudes toward Artificial Intelligence (SATAI) scale. A cross-sectional survey yielded an analytic sample of 672 students at a public teacher education university, randomly divided by academic-year strata into exploratory (n = 334) and confirmatory (n = 338) subsamples. Factor-retention criteria disagreed, and the exploratory solution did not match the component grouping of the source adaptation. A model reproducing that structure fitted poorly and yielded an inadmissible solution. Of seven measurement models estimated from polychoric correlations by diagonally weighted least squares, a bifactor model fitted best (CFI = 0.975; TLI = 0.970), although RMSEA remained 0.086; a four-factor model fitted comparably, but its factor correlations (0.86–0.93) failed the Fornell–Larcker criterion for every pair. Omega hierarchical (0.949) and explained common variance (0.876) indicated a dominant source of common variance, while subscale-specific reliability was limited; internal consistency of the total score was high (α = 0.972). Attitudes were moderately positive (M = 3.59, SD = 0.85), with enjoyment and educational relevance endorsed more strongly than everyday application and self-rated competence. Metric and scalar invariance across academic years was supported by change-in-fit criteria, and fourth-year students scored higher than all earlier cohorts. The findings favor cautious use of the total score rather than subscale scores, and they provide a descriptive basis for designing and then evaluating structured, ethically informed AI preparation in teacher education rather than evidence that such provision is effective.