Automation — including A.I. — accounted for the lion’s share of postwar productivity gains by allowing rapidly improving machines to replace slowly improving human inputs; but essential task complementarities mean any A.I.-led takeoff will be gradual and only becomes explosive once the last bottleneck tasks are automated.

Main Finding

Automation — historically and via modern A.I. — has been the dominant force behind measured TFP growth because it lets rapidly-improving machines replace slowly-improving humans on an expanding set of tasks. However, growth accelerations from A.I. are materially tamed by “weak links”: tasks are complements (elasticity of substitution σ < 1) so output is constrained by the slowest-improving tasks. In calibrated simulations based on historical accounting, A.I. produces a sustained acceleration of growth, but only extreme (computer-sector-like) automation can generate explosive, nearly infinite-income dynamics — and even then the explosion is delayed because of weak links.

Key Points

• Framework
  • Task-based production: a continuum of complementary tasks (σ < 1), each producible by capital or labor with task-specific productivities ψkit (capital) and ψℓit (labor), and an aggregate “other” TFP Zt.
  • Automation = reallocating a task from labor to capital when capital’s (task) productivity relative to labor and factor prices makes capital cheaper.
  • Complementarity (σ < 1) implies “weak links”: aggregate output is highly sensitive to the lowest-output tasks; even infinite output in many tasks does not offset a scarce task.
• Identification / assumptions
  • Key behavioral assumption: tasks automated at a point in time are those with above-average labor costs (equivalently, lower-than-average labor productivity); i.e., automation targets relatively expensive labor tasks.
  • Assumption of no de-automation (once capital replaces labor on a task, it stays replaced).
  • “Heroic measurement”: the set of automated tasks in each year is identified using queries to ChatGPT’s Deep Research (paper notes this as an important caveat).
• Historical accounting results
  • Most TFP growth historically arises from capital- (machine-) augmenting productivity rising much faster than labor-augmenting productivity.
  • Empirical gap: across sectors ψk growth exceeds ψℓ growth by at least ~3 percentage points per year.
  • The residual “other” TFP plus average labor productivity growth (ˆZt + ˆψℓt) is small since 1950 for private business (example: ≈ −0.1%/yr).
  • Counterfactual: freezing the set of automated tasks as of 1950 would eliminate ~87% of TFP growth in private business from 1950–2023 — most gains come from automating more tasks, not simply improving machines on a fixed task set.
• Theory / mechanism
  • Marginal switching (the instant a task is handed to capital because costs equalize) itself produces no productivity gain. The gain from automation is that machines improve faster than humans, so substituting machines for humans on more tasks increases aggregate growth.
  • Because tasks are weak links, aggregate output remains constrained by the slowest tasks until those last bottlenecks are automated.
• Endogenous-growth simulations (two calibrations)
  • “Continuation” (private business calibration): A.I. speeds up automation and idea production modestly; growth accelerates but stabilizes (e.g., ≈2.5% by 2075).
  • “Break”/Computer-sector calibration: if A.I. makes the whole economy behave like the historically fast-computing sector, the model can exhibit explosive (super-exponential) growth; even then the explosion is slow to appear — the model predicts extreme outcomes (infinite income) only after several decades (≈ around 2060 in their extreme calibration).
  • In both cases, there is a positive feedback (“flywheel”): automation → faster idea production → more automation.

Data & Methods

• Data sources
  • Industry- and aggregate-level value-added and factor data from BEA and BLS for the U.S.; sectoral exercises over ~40 years for some industries and ~70 years for aggregate/private business.
• Empirical strategy / growth accounting
  • Start from task-based model with primitives ψkit, ψℓit, Zt and show a reduced-form CES-like representation (At, Bt) that depends on the automation measure βt.
  • Compute growth rates of Bt and At as functions of (i) growth in task-level capital productivity (ψk), (ii) growth in task-level labor productivity (ψℓ), (iii) “other” TFP (Z), and (iv) the flow of automation ˙βt (with task cost shares entering).
  • Identify which tasks are automated each year using ChatGPT Deep Research queries (authors label this a “heroic” measurement choice and discuss robustness).
  • Calibrate the endogenous automation + idea-production model to the historical accounting results (two calibrations: private-business vs computer-sector patterns) and simulate forward.
• Key theoretical results used
  • Proposition: reduced-form CES-like production with capital- and labor-augmenting aggregates Bt, At that inherit dependence on βt.
  • Switching condition: task i uses capital iff ψkit/ψℓit ≥ rt/wt; marginal tasks satisfy equality.
  • Under σ < 1, automation raises Bt by moving more tasks into high ψk mass, but aggregate output remains bounded by the slowest tasks until automation reaches them.

Implications for AI Economics

• Historical interpretation: much of measured postwar TFP growth can be attributed to automation (machines substituting for humans) rather than improvements in human productivity or generic TFP; machines have improved substantially faster than humans at the task level.
• Moderation by bottlenecks: complementarities across tasks (weak links) materially limit the pace at which automation translates into aggregate growth. Even broad, rapid automation leaves aggregate growth constrained by remaining unautomated tasks that improve slowly.
• Two plausible policy-relevant scenarios for A.I.:
  • Continuation-of-history: A.I. accelerates growth modestly and persistently; benefits accrue over decades rather than as an immediate takeoff.
  • Break scenario: A.I.-driven automation that matches computer-industry rates could in principle produce explosive growth, but only after bottleneck tasks are automated — the timing is uncertain and could still be decades out.
• Research and policy priorities implied by the model
  • Identify and target the remaining weak-link tasks: R&D, investment, and training that speed up productivity in tasks that constrain aggregate output will have outsized returns.
  • Measurement: better, transparent, and replicable measurement of which tasks are automated and how ψk and ψℓ evolve is crucial (the paper’s ChatGPT-based task identification is a proof-of-concept, not definitive).
  • Human capital and complementarities: policies improving the productivity of remaining labor-intensive tasks, or easing complementarities (reducing task interdependencies), can materially change growth outcomes.
  • Distributional and transition concerns: while the paper focuses on aggregate TFP, large-scale automation implies shifting factor shares and labor dislocation; policy should prepare for transitions even if aggregate growth remains modest.
• Caveats and uncertainty
  • Identification relies on the assumption that automation targets relatively costly labor tasks and on an operational (and acknowledgedly heroic) task-measurement using ChatGPT; conclusions are sensitive to these choices.
  • The σ < 1 (weak links) assumption is central; if tasks are closer to substitutes, the constraints and timing of acceleration differ markedly.
  • Simulated future paths depend heavily on calibration choices (private-business vs computer-sector patterns) and on structural assumptions about the automation of idea production.

Summary takeaway: The paper provides a theoretically grounded and data-calibrated account showing that automation (including A.I.) has been and can be a dominant driver of growth because machines improve faster than humans. But because production requires many complementary tasks, “weak links” substantially tame and delay any explosive growth from automation — making the future trajectory of A.I.-driven growth hinge on which tasks remain slow and how quickly those bottlenecks are removed.