This study examines the paradox of asymmetric artificial intelligence (AI) use in K-12 education, where teachers embrace AI tools for personal and professional purposes while restricting student access. Drawing on data from the LENS-AI study (N=381) conducted in Baden-Württemberg, Germany, and situating these findings within the broader international research landscape, we identify a consistent pattern across educational systems: teachers act as gatekeepers who benefit from generative AI while protecting students from its use. Our mixed-methods survey reveals that 77% of German teachers use AI privately, yet only 64% employ it for instruction, and 57% completely prohibit student use for ideation tasks. Correlation analyses reveal that both private and educational AI use correlate positively with perceived life improvement and enjoyment, with educational use showing notably stronger associations (ρ=.49 and ρ=.57, respectively) than private use (ρ=.30 and ρ=.33). Teachers using AI in instruction also report greater trust in AI tools (ρ=.44) and less fear (ρ=-.21), suggesting that hands-on professional experience may be particularly effective in fostering positive attitudes. These findings align with international studies from the United States, Australia, China, and Europe, and partially support technology acceptance models, as experience correlates with positive attitudes. However, the persistent gap between teacher and student AI use - despite positive teacher attitudes - suggests that professional identity concerns and risk perceptions focused on cognitive threats (89% of stated concerns) operate as additional barriers beyond TAM constructs. We propose an integrated framework for understanding teacher AI adoption that synthesizes international findings and offers three policy recommendations: professional development addressing values and role perceptions, school development providing legal certainty and spaces for collegial experimentation, and education policy acknowledging cultural and professional dimensions of AI integration.
Autenrieth et al. (Thu,) studied this question.