A two-dimensional view of AI risk: among 664 U.S. occupations, five skill-based clusters show that automation exposure often diverges from job-structure protections, meaning some highly AI-exposed roles are well buffered while other ostensibly safe jobs lack the psychosocial safeguards that mitigate harm.

Main Finding

AI-related occupational risk is multidimensional: automation exposure (measured by an Artificial Intelligence Occupation Exposure — AIOE — score) and job-structure protective features (Psychosocial Buffer Index — PBI) vary independently across occupations. Clustering 664 U.S. occupations on 128 skill dimensions yields five clear occupational groups that differ systematically in both AI exposure and buffering capacity, implying that vulnerability to AI-driven change depends as much on work structure as on automation potential.

Key Points

• Two distinct vulnerability axes:
  • AI Exposure (AIOE): how susceptible a role’s tasks are to AI-driven automation/augmentation.
  • Psychosocial Buffering (PBI): presence of protective job features (e.g., autonomy, task variety, social interaction) linked to occupational health and resilience.
• Data: 664 occupations, 128 skill features (presumably O*NET-style).
• Methods: K-means clustering identified five occupation clusters; Principal Component Analysis (PCA) examined how AIOE and PBI relate to the skill space.
• Results:
  • Five occupational clusters emerged with distinct skill profiles.
  • High-AIOE clusters are heterogeneous in PBI: some high-exposure jobs have strong buffering (lower vulnerability), others have weak buffering (higher vulnerability).
  • Some low-AIOE occupations nonetheless have limited PBI and thus remain vulnerable to stress or negative outcomes.
  • PCA indicates AIOE correlates with cognitive intensity dimensions, while PBI spreads independently across skill components.
• Principal implication: simple “can-this-job-be-automated” metrics are incomplete; job structure matters for economic and health outcomes from AI adoption.

Data & Methods

• Sample:
  • 664 U.S. occupations represented along 128 skill/task dimensions (source likely O*NET or similar).
• Measures:
  • Artificial Intelligence Occupation Exposure (AIOE): a score quantifying susceptibility to AI-related automation/augmentation (constructed from task/skill match to AI capabilities).
  • Psychosocial Buffer Index (PBI): composite index of protective job features (autonomy, task variety, social interaction, etc.) that mitigate occupational stress and help adaptation.
• Analytical approach:
  • K-means clustering (unsupervised) to group occupations by similarity in 128-dimensional skill space; solution with five clusters reported.
  • Principal Component Analysis to map relationships among AIOE, PBI, and underlying skill composition; used to show alignment of AIOE with cognitive-intensity components and the orthogonal dispersion of PBI.
• Key outputs:
  • Cluster profiles showing joint distributions of AIOE and PBI across occupation groups.
  • PCA loadings demonstrating that AIOE is aligned with cognitive task dimensions, while PBI is not simply a byproduct of those same dimensions.
• Limitations (noted or implied):
  • Cross-sectional analysis — cannot capture dynamic adoption or firm-level heterogeneity.
  • AIOE depends on assumptions about AI capabilities and task mappings; results sensitive to how exposure is operationalized.
  • PBI is an index constructed from occupation-level features and may omit workplace-level variation (e.g., managerial practices).
  • K-means and chosen k (five) impose structure; other clustering choices might yield different groupings.
  • Results reflect U.S. occupational structure and may not generalize internationally.

Implications for AI Economics

• Rethinking “automation risk”:
  • Policy and economic models should move beyond binary or scalar automation probabilities and incorporate job-structure features that condition economic and health outcomes.
• Targeting interventions:
  • Prioritize support (retraining, income protection, job redesign) for occupations with the high AIOE + low PBI combination — these workers face the greatest practical and psychosocial vulnerability.
  • Low-AIOE but low-PBI occupations also warrant attention for occupational health interventions even if direct automation risk is small.
• Labor market dynamics and inequality:
  • Heterogeneity in buffering implies uneven adjustment costs and welfare impacts across workers, potentially amplifying inequality if high-exposure low-buffer workers are concentrated in lower-wage groups.
  • Wages, bargaining power, and mobility models should incorporate PBI-like features as moderators of displacement costs and re-employment outcomes.
• Firm strategy and technological design:
  • Employers should orient AI adoption toward augmentation that increases PBI (e.g., tools that raise autonomy, task variety, or collaborative skills), not only productivity.
  • Designing AI as a collaborative partner may reduce negative psychosocial impacts and improve acceptance and productivity gains.
• Measurement and empirical work:
  • Future empirical models of AI impacts should include multidimensional measures (exposure × buffering) to predict outcomes like displacement, wage changes, health, and turnover.
  • Longitudinal and firm-level data are needed to observe how buffering alters realized impacts of AI adoption.
• Policy tools:
  • Combine active labor-market policies (targeted retraining, credentialing) with workplace interventions (job redesign, managerial training) and social insurance for high-risk groups.
  • Use occupational clustering to allocate resources more efficiently: interventions can be cluster-tailored rather than one-size-fits-all.
• Research priorities:
  • Validate AIOE and PBI with longitudinal employment, health, and earnings outcomes.
  • Explore heterogeneity within occupations (firm-level practices) and across countries.
  • Model equilibrium effects: how reallocation, wage adjustments, and firm behavior interact with buffering capacity.

Short summary takeaway: assessing AI-driven labor market risk requires both exposure metrics and measures of job structure that buffer workers. Policy, firm strategy, and economic modeling should incorporate this two-dimensional view to better predict and mitigate distributional and welfare impacts.