To test Huang and Rust’s AI intelligence hierarchy against observed generative AI usage: whether task adoption and the split between automation and augmentation follow the predicted mechanical-to-empathetic order, and how usage moves over time.
This is the first test of the framework against observed generative AI behavior rather than predicted exposure. It speaks to mass-unemployment forecasts built on exposure indices and proposes an intelligence-embodiment matrix that extends the framework toward physical AI.
2,712 O*NET task statements were classified into the four intelligences by a large language model, blind-validated by two human coders, and linked to five waves of the Anthropic Economic Index. Two-part fractional models with cluster-robust and bootstrap inference.
Adoption ignores the hierarchy: physical demands predict non-use, while analytical, intuitive, and empathetic tasks are adopted alike. Engagement forms two tiers: users delegate rule-based work and collaborate on judgment-based work. Empathetic usage grows fastest, driven mainly by consumers.
Two amendments: embodiment governs adoption, and a rule-versus-judgment boundary consistent with output verifiability shapes engagement. The intelligence-embodiment matrix restores the mechanical-first prediction as a testable claim about physical AI. A transparent LLM-coding workflow with blind human validation.
Observed adoption is selective, collaborative, and partly personal, which tempers mass-unemployment forecasts. The matrix shows which tasks face delegation pressure now, which require workflow redesign, and which await physical AI.