Artificial intelligence (AI) exposure, defined as the extent to which occupational tasks can be performed, substituted, or augmented by AI, provides a task-based way to study the environmental consequences of technological change. We construct total, substitution-oriented, and empowerment-oriented AI exposure measures and examine their relationship with industrial carbon emissions using an unbalanced, listed-firm-based province-industry-year panel covering 29 Chinese provinces, 52 industries, and 2016–2023. Baseline fixed-effects estimates show negative associations between AI task exposure and listed-firm carbon emissions, but these estimates are interpreted as conditional associations rather than definitive causal effects. Additional tests using fixed 2016 occupational recruitment weights, province-by-year and industry-by-year fixed effects, and continuous-firm carbon outcomes show that the negative association is most robust for the common intensity of AI task exposure. The mechanism evidence is strongest for industrial upgrading, supportive but less precise for green innovation, and suggestive for energy intensity. The findings support a cautious task-exposure interpretation of AI-related decarbonization: common AI exposure intensity is robustly associated with lower emissions, whereas the substitution-versus-empowerment composition margin is informative but less precisely identified.