Causal Inference with Generative Artificial Intelligence: Application to Texts as Treatments

The paper introduces a GenAI-Powered Inference (GPI) methodology that enhances causal inference with unstructured high-dimensional treatments like text by utilizing large language models (LLMs), specifically Llama 3. This approach leverages the internal representations of LLMs to disentangle treatment features from confounding variables, thus providing more accurate estimates of causal effects without the need for learning causal representations from data. The study includes theoretical foundations for nonparametric identification of average treatment effects and demonstrates the methodology's superiority over existing causal representation learning algorithms through simulation and empirical studies.