Counterfactual design
Workers, workplace geography, and random seed are held fixed so the comparison isolates the policy regime.
Pablo Zavala · AI Safety Evaluation · Research Engineering
Workforce Transitions Under AI Automation
An agent-based NetLogo model of a small labor market adjusting to AI automation. With identical workers, geography, and random seed, peak unemployment reaches 14.3 percent under a tech-driven policy regime versus 3.6 percent under a human-centric one.
14.3 percent vs 3.6 percent peak unemployment under paired policy regimes
Mechanism demonstration in a small simulated labor market, not a macro forecast.
Role: Simulation designer: paired-seed counterfactual, policy regimes, and reproducible benchmark.
How to Inspect This Work
Mechanism visual
The chart shows paired unemployment paths, making the policy-dependent peak visible rather than burying it in a table.
Scope
The result is a mechanism demonstration in a small simulated labor market, not a macroeconomic forecast.
Case Study
Problem
Automation policy debates often treat labor-market impacts as exogenous, even though retraining capacity and adoption speed change the shock itself.
Setup
The NetLogo model compares a tech-driven regime with a human-centric regime while holding workers, geography, and random seed fixed.
Method
A paired-seed counterfactual isolates the policy regime, with local spillovers and a capacity-constrained training system driving the dynamics.
Result
Peak unemployment reaches 14.3 percent under the tech-driven regime and 3.6 percent under the human-centric regime.
Limitation
The model is a small simulated labor market, so the result is a mechanism demonstration rather than a macro forecast.
Evidence
The public repository includes the NetLogo model, seeded benchmarks, and generated figures.
Key Outcomes
- Peak unemployment of 14.3 percent under the tech-driven scenario versus 3.6 percent under the human-centric one
- Identical workers, workplace geography, and random seed; only the policy regime changes
- Seeded, reproducible benchmarks regenerate from committed data
Methods
- Agent-based modeling
- Paired-seed counterfactuals
- NetLogo