Who Adopts Matters: Generative AI and Short-Run Wage Compression

Generative AI is diffusing unevenly across the workforce. Policymakers rely on aggregate productivity estimates to guide resource allocation, yet standard linear aggregation implicitly assumes perfect substitutability between worker types. We embed the micro-data of Bick et al. (2026) in a constant elasticity of substitution (CES) framework (\(\:\sigma\:=1.4\)) to examine distributional consequences that the linear approach cannot capture. Three findings emerge. First, the aggregate productivity estimate is robust: the CES-linear gap is below 0.01 percentage points. Second, asymmetric adoption compresses the skill-service price premium by 0.65%, scaling to approximately 2% under projected adoption. Third, conditional time savings are nearly identical across skill groups (4.5% vs. 4.3% of work hours); the between-group asymmetry operates entirely through differential adoption rates (41.9% vs. 22.9%). These results demonstrate that standard AI impact assessments can obscure distributional consequences when adoption is heterogeneous across worker groups.