This study presents the development of the Belief Scale for Artificial Intelligence in Mathematics Teaching (B-AIMT) and provides psychometric evidence for its initial validation as a domain-specific instrument assessing teachers’ beliefs about AI in mathematics instruction. Based on data from 442 German pre-service and in-service teachers, confirmatory factor analyses provided support for a theoretically grounded multidimensional belief structure comprising nine latent belief dimensions organized within two overarching domains: beliefs about student mathematics learning with AI and beliefs about mathematics teaching with AI. The resulting subdimensions capture perceived instructional affordances, epistemic potentials, perceived risks, and ethical considerations related to AI systems. The final measurement model showed satisfactory model fit and good internal consistency. Measurement invariance analyses supported scalar invariance across pre-service and in-service teachers, allowing meaningful comparisons of latent means. These comparisons revealed generally moderate differences, with pre-service teachers reporting higher latent means on several opportunity- and risk-oriented belief dimensions. Overall, the findings suggest that teachers’ AI-related beliefs form a differentiated yet coherent belief system characterized by a balanced consideration of opportunities and risks. The B-AIMT provides a theoretically grounded instrument that can support research as well as the design and evaluation of AI-related teacher education and professional development in mathematics.