AI beliefs about economic parameters
When you ask a frontier language model for the elasticity of taxable income, it answers with a number and an uncertainty band. This project elicits those answers systematically — 31 models, 26 quantities, 15 independent runs each — and maps where the models agree, where they diverge, and what their answers imply for policy.
Rankings flip by domain
No model family is uniformly more elastic. The most elastic models on labor-and-tax quantities differ from the most elastic on macro-and-trade.
Labor and tax
- Claude Sonnet 4.6
- Grok 4.20
- Grok 4.5
Macro and trade
- Grok 4.3
- Grok 4.20
- GPT-5.6 Luna
Generations converge on core parameters
Across six elicitation waves and ten organizations, pooled centers for the core preference parameters sit in tight cross-model ranges.
- Annual discount factor0.96–0.98
- Coefficient of relative risk aversion1.57–2.10
- Frisch elasticity of labor supply0.28–0.59
A near-zero mean can hide two camps
Gemini 3.5 Flash's capital-gains runs split 10 negative to 5 positive — a bimodal answer that averages to nearly zero without any run saying zero.
The nine headline elasticities
Every panel: 31 models sorted by pooled center, dot at the center, bar spanning the pooled 90 percent interval. Shaded regions are review-based literature ranges. Click through for run-level detail.
Employment participation elasticity of single mothers
Frisch elasticity of labor supply
Income elasticity of labor supply
Uncompensated wage elasticity of labor supply
Capital gains realizations elasticity
Elasticity of taxable income
Intertemporal elasticity of substitution
Elasticity of substitution between capital and labor
Armington elasticity
12,090 successful runs · elicited April through August 2026 · v4 prompts · 15 runs per model-quantity cell. Code · Raw responses · Paper (PDF)