Imagine teaching a powerful AI the nuances of political science. That's the challenge researchers tackled in "Codebook LLMs," exploring whether Large Language Models (LLMs) can interpret complex political texts using specialized guidelines called codebooks. These codebooks, essential tools for political scientists, define concepts like "protest" or "terrorism" with specific criteria and examples. The research asks: can LLMs go beyond simply matching keywords and truly grasp the meaning behind these terms? The team experimented with three real-world datasets – protests in the US, political violence in Pakistan, and political manifestos – comparing how well an LLM understood the concepts with and without detailed codebook instructions. They found that while simply providing labels wasn't enough, giving the LLM the full codebook only offered marginal improvement. Intriguingly, even after training the LLM on specific codebooks (a process called instruction-tuning), the AI struggled to generalize to new, unseen codebooks. This highlights a key limitation: while LLMs excel at pattern recognition, they still struggle with the kind of nuanced understanding that political science demands. This research opens exciting doors for future development. Imagine LLMs helping refine codebooks, identifying inconsistencies, and ultimately speeding up the process of analyzing complex political data. However, it also underscores the need for robust checks to ensure AI accurately reflects the rich tapestry of political meaning.