Calibrated dynamics for China suggest AI capital spurs both investment and wages: AI drives physical-capital accumulation and boosts the wage bill, while feedback from capital and labor onto AI is weak; equilibria are stable rather than cyclic.
Governance of Technological Transition: A Predator-Prey Analysis of AI Capital in China's Economy and Its Policy Implications
semantic_scholar
theoretical
low evidence
7/10 relevance
Full text
usable extracted full text
DOI
Source
PDF
A calibrated three-way Lotka–Volterra model for China (2016–2023) finds AI capital acts like a 'prey' that strongly drives physical capital accumulation and raises labor compensation, with equilibria that are stable and labor outcomes dominated by AI-related parameters.
The rapid integration of Artificial Intelligence (AI) into China's economy presents a classic governance challenge: how to harness its growth potential while managing its disruptive effects on traditional capital and labor markets. This study addresses this policy dilemma by modeling the dynamic interactions between AI capital, physical capital, and labor within a Lotka-Volterra predator-prey framework. Using annual Chinese data (2016-2023), we quantify the interaction strengths, identify stable equilibria, and perform a global sensitivity analysis. Our results reveal a consistent pattern where AI capital acts as the'prey', stimulating both physical capital accumulation and labor compensation (wage bill), while facing only weak constraining feedback. The equilibrium points are stable nodes, indicating a policy-mediated convergence path rather than volatile cycles. Critically, the sensitivity analysis shows that the labor market equilibrium is overwhelmingly driven by AI-related parameters, whereas the physical capital equilibrium is also influenced by its own saturation dynamics. These findings provide a systemic, quantitative basis for policymakers: (1) to calibrate AI promotion policies by recognizing the asymmetric leverage points in capital vs. labor markets; (2) to anticipate and mitigate structural rigidities that may arise from current regulatory settings; and (3) to prioritize interventions that foster complementary growth between AI and traditional economic structures while ensuring broad-base distribution of technological gains.
Summary
Main Finding
The paper models interactions among AI capital, physical capital, and labor in China (2016–2023) with a generalized Lotka–Volterra (predator–prey) framework and finds that AI capital behaves like the “prey”: it stimulates both physical capital and labor (positive b21 coefficients) while receiving only weak negative feedback from them (small-magnitude negative b12). Both subsystems (AI–physical capital and AI–labor) converge to stable node equilibria (no persistent cycles). Global (Sobol) sensitivity analysis shows an asymmetry: labor-market equilibria are predominantly driven by AI-side parameters, whereas physical-capital equilibria are mainly governed by physical-capital self-limiting (saturation) dynamics plus some AI spillovers.
Key Points
- Modeling choice: Uses the Lotka–Volterra predator–prey form to capture nonlinear, bidirectional interactions rather than treating AI as simple exogenous technical change.
- Empirical result (AI ↔ physical capital):
- Continuous-form system:
- dx/dt = 3.852613 x − 0.006965 x^2 − 0.000048 x y
- dy/dt = 4.934909 y − 0.000126 y^2 + 0.007846 x y
- Nontrivial equilibrium: (x, y) = (198.18, 51,506.42)
- Linearization eigenvalues: λ1 = −2.29, λ2 = −5.57 → stable node
- Interpretation: physical capital benefits from AI (b21 > 0); physical capital weakly constrains AI (b12 < 0, small).
- Continuous-form system:
- Empirical result (AI ↔ labor):
- Continuous-form system:
- dx/dt = 3.741844 x − 0.000943 x^2 − 0.000081 x y
- dy/dt = 4.480796 y − 0.000187 y^2 + 0.020083 x y
- Nontrivial equilibrium: (x, y) = (186.78, 44,021.09)
- Eigenvalues: λ1 = −2.52, λ2 = −5.89 → stable node
- Interpretation: AI raises wages/labor totals (b21 > 0); labor exerts only a weak negative feedback on AI.
- Continuous-form system:
- Fit quality: Low MAPE values (AI MAPE ≈ 6.1–6.3%; physical capital ≈ 1.25%; labor ≈ 1.75%) and high adjusted R^2 (≈ 0.99) for the discrete regressions used to estimate parameters.
- Global sensitivity (Sobol; ±10% parameter perturbation, Saltelli sampling N = 1,024):
- For AI equilibrium x*: main drivers were AI intrinsic growth a1 (total-order ST ≈ 0.334), physical-capital self-limiting b22 (ST ≈ 0.254), and the AI→capital coupling b12 (ST ≈ 0.249).
- For physical-capital equilibrium y*: dominated by physical-capital self-limiting b22 (ST ≈ 0.542), then a2 (ST ≈ 0.175) and AI→capital coupling b12 (ST ≈ 0.106).
- For AI–labor subsystem: authors report that labor equilibrium is overwhelmingly driven by AI-side parameters (high Sobol indices on AI growth & AI–labor coupling), i.e., labor outcomes are highly sensitive to AI dynamics.
- Contextual interpretation: In China’s state-led policy context, the estimated weak negative feedbacks on AI are consistent with employment-friendly regulation and restrictions on indiscriminate AI-only subsidies, leading to a predictable, policy-mediated convergence rather than volatile adjustment.
Data & Methods
- Data:
- Annual observations for China, 2016–2023 (n = 8).
- AI capital proxy: AI market size measured by funding raised by AI companies across six AI sub-sectors (AI Robotics, Autonomous & Sensor Technology, Computer Vision, Machine Learning, NLP, Generative AI).
- Physical capital: total investment in fixed assets.
- Labor: total wage bill = average wage of urban employees × urban employed population.
- Sample values are given in the paper (AI capital rises from 15.4 billion yuan in 2016 to 213.7 billion in 2023; physical capital and labor series also reported).
- Model and estimation:
- Starts with continuous Lotka–Volterra ODEs; converts to a discrete form suitable for annual data (Leslie-style transformation).
- Regression form (zero-intercept method) used to estimate transformed parameters; back-transformed to recover continuous a and b coefficients.
- Goodness-of-fit assessed by adjusted R^2 and MAPE.
- Stability and equilibrium:
- Analytical equilibrium expressions derived; Jacobian evaluated at equilibria; eigenvalues computed to classify local stability.
- Global sensitivity:
- Sobol variance decomposition (first-order and total-order indices) via Saltelli sampling (N = 1,024 base) applied to six model parameters (a1, b11, b12, a2, b21, b22), each perturbed uniformly ±10% around baseline estimates. Nonphysical samples (negative/non-finite equilibria) were rejected.
Implications for AI Economics
- Policy leverage is asymmetric:
- Labor-market outcomes are highly sensitive to AI-side parameters (AI growth and AI→labor coupling). Small shifts in AI promotion or regulation can have outsized effects on wages and total labor compensation; hence labor-targeted policies (retraining, wage supports, social insurance) should be coordinated tightly with AI promotion.
- Physical capital equilibria are largely governed by internal saturation (self-limiting effects). Policies to expand traditional capital (investment incentives, lowering bottlenecks to absorption) are needed if the objective is to raise physical-capital capacity beyond its intrinsic limits; AI spillovers help but are secondary.
- Predictability vs. rigidity:
- Stable-node equilibria indicate current policy settings produce monotonic, predictable convergence rather than oscillations. This is useful for planning but may create structural rigidity that reduces flexibility to reallocate resources optimally in the long run. Policymakers should monitor whether stability reflects desirable complementarity or institutional constraints that lock in suboptimal patterns.
- Targeted interventions:
- Favor policies that promote complementarity (AI that augments existing capital and labor) rather than purely AI-centric resource concentration. Examples: subsidize AI applications that raise productivity in manufacturing or services; incentivize co-investment across tech and traditional firms; design procurement and regional policies to encourage diffusion.
- Because labor equilibria are AI-sensitive, prioritize worker transition policies: training programs aligned with AI-complementary tasks, wage floor measures, and policies that ensure AI benefits are widely distributed.
- Monitor and adjust interaction coefficients in real time: collect better measures of AI capital, measure AI–sector linkages, and use sensitivity-informed indicators (e.g., estimated b12 and b21) to assess whether interventions are steering the system toward desired equilibria.
- Cautions and research needs:
- Short sample and proxy limitations: the analysis uses 8 annual observations and funds-as-capital proxy for AI; results should be interpreted as preliminary and policy-guiding rather than definitive.
- Aggregation and functional form: Lotka–Volterra imposes bilinear interaction structure and aggregates heterogeneous sectors and occupations. Future work should test robustness to alternative specifications (task-based models, sectoral disaggregation, endogeneity controls).
- Endogeneity and policy feedback: state interventions both shape and respond to dynamics; causal identification (e.g., instrumental variables, policy shocks) would strengthen inference on which interventions causally change b12/b21 or a1/a2.
- Operational recommendation for policymakers:
- Use the asymmetric sensitivity results to prioritize (1) monitoring AI growth parameters and AI→labor coupling, (2) policies that enhance AI–capital complementarity while easing physical-capital saturation constraints, and (3) active labor-market measures to capture AI’s gains broadly.
If you want, I can (a) produce a short table summarizing the estimated parameters and equilibria for quick reference, (b) outline specific policy instruments matched to the sensitivity indices, or (c) sketch robustness checks (alternative datasets, longer panels, instrument strategies) you could pursue.
Assessment
Paper Typetheoretical
Evidence Strengthlow — Results are based on calibration of a stylized Lotka–Volterra system to an 8-year aggregate (2016–2023) Chinese time series without exogenous variation or causal identification; short sample, aggregate proxies, and model structure drive conclusions, so empirical support for causal claims is weak.
Methods Rigormedium — The paper uses internally consistent dynamical-systems methods (equilibrium computation, Jacobian/stability analysis, and variance-based global sensitivity analysis), but estimation/calibration on a very short national aggregate panel, limited robustness checks to alternative model specifications, and absence of micro-level validation reduce overall methodological rigor.
SampleAnnual national-level Chinese aggregate series (2016–2023) for three stock-like variables: a proxy for AI capital, physical capital stock, and labor compensation (wage bill); no firm- or sector-level heterogeneity, short 8-year panel.
Themeslabor_markets productivity
GeneralizabilityChina-only national aggregates (may not hold for other countries), Very short time span (2016–2023) limits inference about longer-run dynamics, Annual frequency may miss higher-frequency adjustment or transitional cycles, Aggregate/national-level analysis masks firm-, sector-, and regional heterogeneity, Stylized Lotka–Volterra structure may not capture important economic mechanisms or omitted variables, AI capital is proxied and may suffer measurement error or definition sensitivity, No identification of exogenous policy or technology shocks limits causal generalization
Claims (14)
| Claim | Direction | Outcome | Confidence & Evidence | Details |
|---|---|---|---|---|
| The paper models interactions among AI capital, physical capital, and labor using a Lotka–Volterra (predator–prey type) system adapted to include self-limiting (saturation) terms. Other | null_result | model structure / interaction specification (no single dependent variable) |
Reading fidelity
high
Study strength
low
|
Model structure: Lotka–Volterra–type interactions among AI capital, physical capital, and labor (methodological claim)
|
| The empirical analysis uses annual, national-level aggregate Chinese series for 2016–2023 as proxies for AI capital, physical capital stock, and labor compensation (wage bill). Other | null_result | AI capital proxy; physical capital stock; labor compensation (wage bill) |
Reading fidelity
high
Study strength
low
|
n=8
National annual aggregates for China 2016–2023 used as proxies for AI capital, physical capital, and labor compensation (data description)
|
| Estimated interaction coefficients indicate AI capital positively drives physical capital accumulation (AI → physical capital positive effect). Firm Productivity | positive | physical capital stock / accumulation |
Reading fidelity
medium
Study strength
low
|
n=8
Estimated AI → physical capital interaction coefficient is positive
|
| Estimated interaction coefficients indicate AI capital increases labor compensation (AI → wage bill positive effect). Wages | positive | labor compensation (wage bill) |
Reading fidelity
medium
Study strength
low
|
n=8
Estimated AI → labor compensation (wage bill) coefficient is positive
|
| Feedback effects from physical capital and labor onto AI capital are weak, with only weak negative feedback observed (physical capital → AI and labor → AI small/weakly negative coefficients). Other | negative | AI capital growth/stock (feedback strength) |
Reading fidelity
medium
Study strength
low
|
n=8
Feedbacks from physical capital and labor onto AI are small and weakly negative in the estimated system
|
| Equilibrium points of the estimated three-stock system are classified as stable nodes (no persistent endogenous cycles under the estimated parameters). Other | null_result | equilibrium stability classification (eigenvalues of Jacobian) |
Reading fidelity
medium
Study strength
low
|
n=8
Estimated system equilibria classified as stable nodes (local stability via Jacobian eigenvalues)
|
| Under the current estimated parameters, dynamics converge toward equilibria—implying convergent, policy-mediated adjustment rather than endogenous cyclical instability. Other | null_result | convergence behavior of model trajectories (toward equilibrium) |
Reading fidelity
medium
Study strength
low
|
n=8
Model dynamics converge toward equilibria under estimated parameters (implied convergent adjustment)
|
| Global sensitivity (variance-based) analysis shows labor-market equilibrium outcomes are overwhelmingly driven by AI-related parameters. Wages | positive | labor-market equilibrium (wage bill / labor stock) |
Reading fidelity
medium
Study strength
low
|
n=8
Global sensitivity: labor-market equilibrium outcomes are predominantly driven by AI-related parameters (majority of variance attributed to AI parameters)
|
| Global sensitivity analysis shows physical-capital equilibrium outcomes are jointly influenced by AI–physical interactions and by physical-capital self-limitation (saturation) dynamics. Firm Productivity | mixed | physical capital equilibrium (physical capital stock) |
Reading fidelity
medium
Study strength
low
|
n=8
Global sensitivity: physical-capital equilibrium influenced jointly by AI–physical interactions and physical-capital saturation dynamics
|
| Policy leverage is asymmetric: interventions targeting AI-related parameters have large effects on labor outcomes and nontrivial effects on capital, whereas interventions targeting physical-capital parameters have more limited effects on labor. Wages | mixed | labor compensation (wage bill) and physical capital stock responses to parameter changes |
Reading fidelity
low
Study strength
low
|
n=8
Policy leverage asymmetric: interventions on AI parameters have large effects on labor outcomes; interventions on physical-capital parameters have limited labor effects (model-based)
|
| Promoting AI without complementary policies for physical capital and labor may produce uneven outcomes; policy sequencing and complementarity (capital modernization, workforce upskilling) are recommended to produce more inclusive growth. Governance And Regulation | mixed | distributional and growth outcomes (qualitative policy impacts inferred from model) |
Reading fidelity
speculative
Study strength
low
|
n=8
Promoting AI without complementary physical-capital and labor policies may produce uneven outcomes; sequencing/complementarity recommended (policy implication)
|
| Because feedbacks from capital and labor onto AI are weak, AI can grow rapidly and may lead to lock-in, concentration, and distributional risks that warrant monitoring and possible redistributive or competition policies. Market Structure | negative | AI capital growth dynamics and potential long-run concentration/lock-in risks (qualitative) |
Reading fidelity
low
Study strength
low
|
n=8
Weak feedbacks to AI imply potential for rapid AI growth, lock-in, concentration and distributional risks (qualitative risk implication)
|
| The main empirical conclusions are based on a short annual panel (2016–2023) and a stylized aggregate interaction model; results should be interpreted with caution due to potential omitted variables, aggregation bias, and limited sample size. Other | null_result | validity/robustness of empirical conclusions (limitations) |
Reading fidelity
high
Study strength
low
|
n=8
Main empirical conclusions hinge on a short annual panel (2016–2023) and stylized aggregate model; results subject to omitted variables, aggregation bias, and small sample caveats
|
| Estimation/calibration, stability assessment, and global sensitivity methods used: parameters calibrated/estimated on 2016–2023 data; equilibrium located; Jacobian eigenvalues computed for local stability; variance-based global sensitivity analysis performed over parameter space. Other | null_result | methodological procedures applied (estimation, stability analysis, GSA) |
Reading fidelity
high
Study strength
low
|
n=8
Methods used: parameter estimation/calibration on 2016–2023 data, equilibrium location, Jacobian eigenvalue stability assessment, variance-based global sensitivity analysis
|
Entities
AI capital (ai_tool)
Physical capital stock (outcome)
Labor compensation (wage bill) (outcome)
Annual Chinese national aggregates (2016–2023) (dataset)
Lotka-Volterra (predator–prey) model (deterministic, three-way) (method)
Interaction coefficients (Lotka-Volterra parameters) (method)
Self-limitation (saturation) parameters (method)
Stability analysis (Jacobian eigenvalues) (method)
Global sensitivity analysis (variance-based) (method)
Equilibrium points (steady states) (outcome)
Labor-market equilibrium (outcome)
Physical-capital equilibrium (outcome)
Employment and wage distribution (outcome)
AI–physical-capital complementarities (outcome)
Policymakers (population)