An Explainable AI Approach to Large Language Model Assisted Causal Model Auditing and Development

LLMs can help identify errors in automatically-generated causal models: The paper demonstrates that ChatGPT can audit causal networks inferred from data by examining individual relationships and flagging potential problems like incorrect edge directionality, missing confounders, and unmeasured mediating variables—addressing a key limitation where algorithmic inference often produces "erroneous edges."

Human expertise remains essential in AI-assisted causal reasoning: Rather than replacing domain experts, the proposed approach positions LLMs as one tool in a collaborative team that includes "the human analyst and domain expert," emphasizing that AI should augment rather than substitute for specialized knowledge in fields like medicine, epidemiology, and social science.

Explainability enables better human oversight of AI recommendations: The method uses visualizations to summarize ChatGPT's reasoning about causal relationships, allowing analysts to understand why the AI flagged certain edges and make informed decisions about whether to modify models, collect more data, or test hypotheses—supporting more transparent AI deployment.

This approach is still in early stages: The authors describe their work as "first results obtained with an emerging prototype," indicating this is exploratory research that would require further validation before practical implementation in high-stakes domains like medicine or public health.