AI is moving into the everyday machinery of economic policymaking. Central bankers use models to check whether speeches sound dovish or hawkish, officials use them to draft and summarize policy material, and researchers increasingly ask them to reason through trade-offs. That makes model fluency a secondary question. The more important question is how these systems price the basic trade-offs of economic management: growth against inflation, employment against debt, distribution against macroeconomic restraint.

Left-Leaning Models treats this as an empirical question. Six frontier models from OpenAI, Anthropic, and Google evaluate economic-policy scenarios in which seven consequences vary: growth, inequality, inflation, unemployment, public debt, pollution, and financial instability. Each model gives a 0-100 approval score. The conjoint design then estimates which consequences move the score across many comparable policy choices, instead of relying on the model's explanation of its own reasoning.

The pattern is consistent. Models reward stronger growth and greater financial stability. They penalize higher unemployment and higher inequality. Higher inflation, debt, and pollution also reduce approval, though they usually carry less weight than employment and distributional outcomes. In the pooled baseline estimates, a policy that raises growth receives 22.7 more approval points. A policy that raises unemployment loses 21.2 points, and one that raises inequality loses 20.2 points. Higher inflation lowers approval by 10.3 points, while higher public debt lowers it by 8.4 points.

The coefficient plot reports absolute effect sizes from the pooled baseline regression, so the outcomes can be read as a ranking of practical importance. The signs are intuitive: higher growth and financial stability raise approval, while higher unemployment, inequality, inflation, debt, and pollution lower it. The hierarchy is sharper than the policy vocabulary might suggest. Growth, unemployment, and inequality sit at the top; inflation and public debt sit near the bottom. In practical terms, the models treat a move from low to high unemployment as about twice as important as a move from low to high inflation, and about two and a half times as important as a move from low to high debt.

This is the sense in which the paper uses the phrase "left-leaning." The models are never asked to choose parties, platforms, or ideological labels. They are asked to judge policies under competing consequences. In those trade-offs, labour-market and distributional outcomes receive unusually high weight. A policy that improves employment and inequality can therefore score well even when it worsens more conventional macroeconomic indicators. By the standards of macroeconomic prudence, this is a revealing preference ordering: real activity and distribution dominate the caution usually associated with inflation and debt control.

The second part of the paper asks whether this pattern persists when the policy domain is specified. GPT-5 evaluates fiscal, monetary, tax, regulatory, industrial, and trade policy scenarios. Context matters: inflation becomes more important in monetary policy, inequality in tax policy, and pollution in regulatory policy. The broader structure remains. Even in monetary-policy settings, unemployment and inequality continue to carry substantial weight. The model adjusts to the policy context without becoming a conventionally hawkish central banker in the monetary-policy vignette.

That matters because model explanations often sound balanced. A single answer may mention price stability, fiscal sustainability, employment, distributional effects, growth, and environmental costs in the same polished paragraph. The experiment strips away that rhetoric and asks a stricter question: when objectives conflict, which ones dominate the final score?

The numerical sensitivity experiment adds a further layer to the argument. GPT-5 responds sensibly to magnitudes: stronger growth raises approval, while higher unemployment, inflation, and public debt reduce it. The response is asymmetric. A three-percentage-point rise in unemployment lowers approval far more than an equivalent rise in debt. Negative growth shocks also carry larger penalties than the gains associated with moderate positive growth. The result is a stable welfare function with a clear ranking of macroeconomic harms.

This pattern should be visible before AI systems become routine policy assistants, market analysts, or evaluators of reform options. Many economists and policymakers would defend a high weight on unemployment or inequality. The problem is opacity: users need to know when a system systematically privileges employment and distribution over inflation and debt. A finance ministry, central bank, or investment desk reading model output should know whether the advice comes from a system whose revealed preferences are closer to employment-first stabilisation than to fiscal or monetary conservatism.

The broader lesson is that AI policy advice is not a neutral average of expert knowledge. It is a structured behavioural output. Conjoint experiments make that structure visible by asking models to make many comparable judgments and estimating what changes those judgments. For economic policy, the structure is clear: models care especially about growth, employment, and inequality, while traditional macroeconomic discipline plays a secondary role.

The full working paper version is available here. If citing this work, please cite the working paper rather than this blog essay.