Generative AI won't mechanically eliminate work — its economic impact depends on institutions: adoption, verification, apprenticeship and bargaining shape who benefits and whether expertise endures. Policymakers and firms must treat AI deployment as governance and workflow design, not plug-and-play automation.

Main Finding

The paper argues that asking whether AI will "destroy jobs" is the wrong question. Instead, we need a task-based, institutionally grounded framework: generative AI should be seen as cheap, scalable, but fallible cognition. Labor-market outcomes depend not on a single technical metric but on exposure, adoption, verification, workflow design, apprenticeship, and institutional bargaining. AI changes where scarce human capital is valuable (upstream question selection, framing, interpretation, responsibility), and its net effect on employment and expert formation is an institutional equilibrium shaped by governance, liability, and incentives — not an automatic slide to mass unemployment or guaranteed abundance.

Key Points

• Jobs are bundles of tasks and governance arrangements, not indivisible units; machine cognition is heterogeneous in what it can substitute.
• Introduces an eight-margin taxonomy that determines AI’s labor effects:
  • Exposure (which tasks are technically reachable)
  • Adoption (when firms actually use AI for a task)
  • Verification (costs to check AI outputs)
  • Question selection (who asks the economically meaningful questions)
  • Workflow redesign (how tasks are recombined)
  • Demand elasticity (how cheaper cognition affects demand for services)
  • Apprenticeship (how junior tasks train future experts)
  • Rent allocation (who captures gains from cheap cognition)
• Distinguishes technical reach (what models can do) from equilibrium displacement (what happens in markets once institutions and frictions are considered).
  • Proposes a task vulnerability index to quantify technical exposure.
  • Proposes an adoption condition that makes verification, liability, trust, and governance explicit in the decision to deploy AI.
• Reframes occupations as governance bundles (bundles of rights, responsibilities, verification practices) rather than simple task lists; firms are architectures of distributed intelligence coordinating human and machine cognition.
• Labor-market effects arise from a balance of:
  • Task compression (automation reduces human steps)
  • Scale expansion (lower marginal costs expand output)
  • Creation of new human work (supervision, interpretation, upstream tasks)
  • Institutional bargaining over rents (who benefits)
• Central dynamic: expertise formation. Many junior tasks simultaneously produce current output and train future judgment; automating them can boost short-run productivity but weaken the pipeline to accountable expertise unless AI is designed to teach and scaffold learning.

Data & Methods

• Conceptual and theoretical development rather than primary empirical estimation.
• Tools used:
  • Task-based microeconomic modeling to separate technical exposure from adoption equilibrium.
  • Definition of a task vulnerability index (technical reach metric) and an explicit adoption condition incorporating verification and governance costs.
  • Institutional modeling that treats occupations as governance bundles and firms as coordination architectures.
  • Qualitative analysis of margins (exposure, verification, apprenticeship, etc.) and equilibrium channels (task compression, scale expansion, demand effects, rent allocation).
• Emphasis on linking technical characteristics of generative models (cheap, scalable, fallible cognition) with organizational, legal, and market frictions that determine real-world outcomes.
• Note: the paper is primarily a theoretical and descriptive synthesis; empirical validation is left as a research agenda.

Implications for AI Economics

• Measurement and empirical priorities:
  • Move beyond occupation-level automation probabilities to task-level vulnerability indices, measures of verification costs, and metrics of apprenticeship loss.
  • Track adoption conditional on governance, liability, and trust factors, not just model performance.
• Policy and institutional design levers:
  • Invest in apprenticeship and training systems that preserve pathways to accountable expertise; design AIs to teach/junior-upskill rather than simply bypass.
  • Clarify liability, verification, and governance rules to shape adoption incentives (e.g., standards for verification, auditing requirements).
  • Consider competition and rent-allocation policies: cheap cognition can create large rents; policy choices determine who captures them (firms, workers, consumers).
  • Support worker voice and bargaining institutions to influence workflow redesign and distributional outcomes.
• Firm and organizational strategy:
  • Firms should treat AI deployment as workflow and governance redesign, not plug-and-play automation; anticipate shifts in scarce human capital toward question formation, framing, interpretation, and responsibility.
  • Design architectures of distributed intelligence to preserve expertise pipelines (e.g., human-in-the-loop training, teaching-oriented AI tools).
• Macroe and labor-market outlook:
  • Neither inevitable mass unemployment nor guaranteed abundance follows automatically from AI progress; outcomes are institutional equilibria.
  • Short-run productivity gains from automating junior tasks can coexist with longer-run weakening of expert supply unless apprenticeship and learning are preserved.
• Research agenda:
  • Empirically estimate verification costs and their effect on adoption.
  • Quantify how automating junior tasks affects long-run expert formation and labor supply of high-responsibility roles.
  • Study how different liability and governance regimes change adoption, rents, and distributional consequences.