Evidence-based policy or beauty contest? An LLM-based meta-analysis of EU cohesion policy evaluations

Independent and high-quality evaluations of government policies are an important input for designing evidence-based policy. Institutional frictions and lack of incentives to write such evaluations, on the other hand, carry the risk of turning the system into a costly beauty contest. We study one of the most advanced markets of policy evaluations in the world, the evaluations of EU Cohesion Policy interventions by the EU Member States. We use a large language model to quantify the findings of about 2,300 evaluations, and complement this data with our own survey of the evaluation authors. We show that the findings of evaluations are inconsistent with those of the academic literature on the output impacts of Cohesion Policy. Using further variation across Member States, our analysis suggests that the market of evaluations is rather oligopolistic within Member States, that it is very fragmented across the EU, and that there is often a strong involvement of managing authorities in the work of formally independent evaluators. These factors contribute to making the findings of the evaluations overly optimistic (beautiful) risking their relevance for (evidence-based) policy. We conclude by discussing reform options to make the evaluations of EU Cohesion Policy more unbiased and effective.