Transformative AI could rewrite long-run growth: when AI both boosts productivity and autonomously accumulates capital, self-reinforcing feedbacks can permanently raise growth or destabilize equilibria depending on complementarities and absorptive capacity.

Main Finding

The paper extends Solow (exogenous) and Romer (endogenous) growth frameworks to treat Transformative AI (TAI) as an autonomous, recursive driver of technological progress. By adding AI-driven feedback terms to total factor productivity (TFP) and innovation dynamics, the author argues that TAI can produce non-linear, super- or hyper-exponential growth, undermine the Solow steady-state/convergence logic, and generate persistent divergence, new distributional pressures, and novel policy/governance challenges.

Key Points

• Conceptual contribution
  • Treats TAI as a hybrid factor that functions simultaneously as capital, labour substitute, and autonomous innovator.
  • Emphasizes recursive self-improvement: AI raises its own rate of productivity growth through feedback loops (data → better models → more data/value).
• Modified Solow specification
  • Replaces constant exogenous TFP growth with an AI-amplified term: A(t) = A0 · exp[(g + λ · φ(A)) t]
  • φ(A) is a recursive intelligence function (examples: ln(A) or A^θ). If λ·φ(A) becomes large, growth can become super-exponential (hypergrowth/singularity).
  • Implication: diminishing returns to capital need no longer enforce convergence; early adopters gain compounding advantages.
• Modified Romer/endogenous specification
  • Adds an autonomous AI-driven innovation term: Ȧ = δ H_A A^α + η A^β
  • The first term is human-driven R&D; the second captures AI self-improvement. If β > α (and η is large), AI can dominate innovation and produce runaway growth dynamics.
• Dynamics and equilibria
  • Recursive AI terms introduce strong non-linearities, multiple equilibria, potential instability, and absence of classical steady states unless bounded by exogenous/regulatory constraints.
• Distributional, institutional and ecological implications
  • Amplifies global and within-country inequality via concentration of data/computational resources.
  • Blurs labour–capital boundary (“synthetic labour”), potentially decoupling wages from productivity.
  • Raises energy/resource constraints (AI is computationally intensive), so sustainability must be modelled endogenously.
• Policy and governance
  • Argues for algorithmic macroeconomics: policy instruments targeting λ and η (investment, regulation, data governance, ownership rules).
  • Emphasizes need for controls to avoid runaway innovation cycles, address monopoly capture of AI knowledge flows, and integrate environmental constraints.
• Suggested research directions
  • Formal multi-sector dynamic models with feedback learning parameters, agent-based simulations, calibration/empirical estimation of λ, η, β, and studies of stabilizing policy interventions.

Data & Methods

• Nature of contribution: theoretical/modeling extension (no new micro or macro empirical dataset provided).
• Methods used:
  • Analytical modification of canonical production/innovation equations (Solow Cobb–Douglas + exogenous TFP; Romer endogenous knowledge production).
  • Introduction of parametric recursive functions (φ(A), power-law feedback) to capture AI self-improvement.
  • Cites related empirical/agent-based studies and prior literature (e.g., agent-based Solow simulations, empirical measures of AI contributions to TFP) but does not present original simulations or empirical estimation in the paper.
• Limitations in methods (noted or implied):
  • Functional forms (φ(A), parameter values λ, η, β) are illustrative rather than estimated or calibrated.
  • Stability analyses and formal proofs of multiplicity/instability are sketched intuitively rather than derived in full generality.
  • No empirical validation or quantitative forecasting; policy suggestions are conceptual.

Implications for AI Economics

• Modeling: Growth theory must incorporate recursive algorithmic feedback as an endogenous force—introducing new state variables (algorithmic capacity, data stocks, energy costs) and possibly multi-sector/agent-based frameworks to capture non-linear dynamics.
• Measurement: Research agenda should estimate AI-amplification parameters (λ, η) and recursive elasticity (β, θ), quantify AI’s contribution to TFP and to “effective labour,” and measure data and compute concentration across countries/sectors.
• Distribution & development: Standard convergence results may fail; policy must address digital infrastructure and data-access inequality to avoid amplified global divergence and platform monopolies.
• Labour & institutions: Re-define “effective labour” to include algorithmic capacity; reassess education, social insurance, and taxation in light of synthetic labour and potentially decoupled wage-productivity dynamics.
• Governance & macroprudential policy: Design instruments to manage AI-driven instability (controls on recursive AI deployment, data/knowledge ownership rules, algorithmic oversight), and incorporate algorithmic risks into fiscal/monetary/financial stability frameworks.
• Sustainability: Endogenize environmental/exergy constraints (computational energy use) and incentivize “green AI” to align growth with ecological limits.
• Future research priorities: formal dynamic and multi-sector models, agent-based simulations, calibration/empirical testing, policy counterfactuals for bounding η/λ, and institutional analyses of knowledge ownership and international spillovers.

Overall, the paper proposes a conceptual shift: AI transforms technological change from largely human-driven, well-behaved growth processes into potentially self-accelerating, non-linear processes that require new modeling tools, measurement strategies, and policy frameworks in AI economics.