Causal Representation Learning with Generative Artificial Intelligence: Application to Texts as Treatments

Causal Representation Learning with Generative ArtificialIntelligence: Application to Texts as Treatments∗This academic paper explores a novel approach to causal inference with unstructured data like text, focusing on how generative AI, specifically Large Language Models (LLMs), can improve the process. The core idea is to leverage the internal representation of text generated by LLMs to disentangle treatment features of interest from confounding features. The authors propose a method based on a neural network architecture and double machine learning to estimate average treatment effects and extend it to address the challenge of perceived treatment features using an instrumental variables approach. Through simulations and an empirical study using candidate biographies, the paper demonstrates the proposed methodology's effectiveness in reducing bias and improving computational efficiency compared to existing techniques.