Large language models for causal inference

Causal inference provides essential tools for understanding how interventions influence outcomes, supporting evidence-based decision-making across disciplines such as medicine, economics, and the social sciences. At the same time, large language models (LLMs) have rapidly advanced, offering capabilities in text generation, coding, and knowledge synthesis. This thesis examines the overlap of these two areas by looking at the uses of LLMs in causal inference. The systematic review was based on PRISMA principles and included the 2017 to 2024 publications in Scopus, arXiv, PubMed, and the ACL Anthology. Eighty studies were enclosed and examined under four classifications namely causal discovery, extracting causeeffect relations, experimental design support, and causal code generation. A GPT-based assistant prototype was also created in order to test systematic review workflows with the help of AI. The results suggest that LLMs might facilitate applied researchers, making it faster to do things like DAG generation, hypothesis generation, and apply causal inference libraries. Nonetheless, interpretability, robustness, generalization, and reproducibility is still a challenge. The assistant enhanced the efficiency in screening and synthesis although it also brought to the fore risks of model dependency and human bias. All in all, the research provides the answer: LLMs can be a strong tool to use as an augmented research assistant in the process of causal analysis, but only when their claims are justified by transparent and reproducible practice.