New method uses genetic algorithms and LLMs to simulate AI policy outcomes

In a paper accepted at the ACM Conference on Fairness, Accountability, and Transparency (FAccT) 2026, researchers Julia Barnett, Kimon Kieslich, Natali Helberger, and Nicholas Diakopoulos propose a groundbreaking method to evaluate AI policies through large-scale simulation of interventions. The methodology integrates three components: participatory evaluation (surveying citizens on policy preferences), expert assessment of implementation costs, and an LLM-based evaluation of each policy's perceived harm mitigation. A genetic algorithm then runs thousands of simulations, exploring a vast solution space of potential policy combinations (e.g., transparency rules, licensing requirements, audit mandates) and examines how outcomes shift under different weightings of cost, participatory input, and harm mitigation.

The key finding is that the genetic algorithm generates a diverse set of viable policy packages, which can serve as a starting point for deliberation among policymakers. The method operationalizes participatory AI by embedding it directly into practical policy development pipelines. According to the authors, this approach allows policymakers and researchers to assess how much weight to assign to expert versus public input, and to target areas that warrant greater time and resource investment. The study concludes that such simulations could help prioritize among competing AI regulations globally, especially as the rapid proliferation of AI systems and harms outpaces traditional policy assessment.