Scaling Political Texts with Large Language Models: Asking a Chatbot Might Be All You Need

We use instruction-tuned LLMs such as GPT-4, MiXtral, and Llama 3 to position political texts within policy and ideological spaces. We directly ask the LLMs where a text document or its author stand on the focal policy dimension. We illustrate and validate the approach by scaling British party manifestos on the economic, social, and immigration policy dimensions; speeches from a European Parliament debate in 10 languages on the anti- to pro-subsidy dimension; Senators of the 117th US Congress based on their tweets on the left-right ideological spectrum; and tweets published by US Representatives and Senators after the training cutoff date of GPT-4. The correlation between the position estimates obtained with the best LLMs and benchmarks based on coding by experts, crowdworkers or roll call votes exceeds .90. This training-free approach also outperforms supervised classifiers trained on large amounts of data. Using instruction-tuned LLMs to scale texts in policy and ideological spaces is fast, cost-efficient, reliable, and reproducible (in the case of open LLMs) even if the texts are short and written in different languages. We conclude with cautionary notes about the need for empirical validation.