Reliability of Commercial LLMs as Political Advisory Companions

Determine whether commercial large language models can function as reliable political advisory companions for legislators in real-world governance contexts, providing dependable analysis and guidance for legislative decision-making.

Background

Legislators operate under bounded rationality, facing high information loads with limited cognitive and time resources, and increasingly consider LLMs as potential advisory tools. Despite extensive research on political bias and simulation capabilities of LLMs, there is limited evidence on whether these systems can be trusted for day-to-day legislative advisory work.

The paper frames the legislator–LLM relationship using principal–agent theory and evaluates multiple commercial and open-weight models against official Romanian legislative explanatory memoranda. The explicit open question concerns the fundamental reliability of commercial LLMs as political companions within actual parliamentary workflows.