Effectiveness of LLM Election Safeguards and Extent of Developer Control

Determine the practical effectiveness of election-related safeguards implemented in large language model platforms, characterize how widely deployed large language models behave in election-related interactions, and ascertain the extent to which developers can control large language model behavior during electoral contexts.

Background

The paper studies how LLMs behave during the 2024 US presidential election season and notes that major providers have introduced safeguards to prevent misuse or unintended influence. Despite these measures, the authors highlight that it is not established whether such safeguards are effective in practice, what the real-world behavior of LLMs looks like in election contexts, or how much control developers ultimately have over model behavior.

This uncertainty motivates the authors’ longitudinal evaluation of multiple models over time, designed to capture systematic variation, steering sensitivity, and guardrail effects. The quoted sentence explicitly frames these as open questions in need of rigorous answers.