← all posts

ai already runs policy simulation

i asked whether ai has been used for policy making, imagining a future where we simulate every policy change on a giant model of citizens and vote on the best one, with humans still deciding. it exists, it is running, and it is much narrower than that. policyengine is the real answer, open source, tens of thousands of policy parameters encoded, used by the uk government and us congressional offices, and it runs microsimulation on real tax data, so you can install it and score a policy reform in minutes.

there are bigger national systems too. one country runs a national policy simulation platform with over a thousand analytical projects across government entities and claims major identified savings plus a road fatality rate cut using machine learning on traffic data. a research institute shipped an agentic policy tool that models social and economic impact before implementation, and a commercial product runs twenty agents across domains like economist, trade advisor, and fiscal analyst, sold explicitly to policymakers.

but the full version i wanted does not exist because it cannot. llm agents pretending to be households simulate the model's opinions, not people, and there is no objective answer to who should win when a policy helps one group and hurts another. picking those weights is the political decision, so simulation gives you who gains and who loses, and a human still has to pick. any system claiming to output the optimal policy has hidden its value judgments somewhere.