Generative Models and Causality – Kyungwoo Song

This seminar examines how generative AI advances three foundational tasks in causality, treated as distinct, modular problems: (1) causal inference via intervention‑effect estimation, (2) causal graph analysis, and (3) detection of causal mechanism shifts and change points. First, for causal inference, we consider procedures in which generative models align domain knowledge with observational signals to represent treatment, confounding, and temporal context. This enables stable estimation of intervention effects and principled policy evaluation without relying on explicit counterfactual generation. Second, for causal graph analysis, we outline strategies that combine language‑grounded knowledge extraction and constraint proposals with statistical checks to improve the reliability of directionality and structure, yielding interpretable hypothesis spaces and testable causal claims. Third, for shift detection, we describe methods that disentangle changes in functional mechanisms from changes in noise, supporting early diagnosis of performance degradation, targeting of monitoring resources, and evidence‑based model updates in deployed settings. Across these tasks, generative AI serves as a computational aide for knowledge alignment, hypothesis proposal and pruning, uncertainty annotation, and experiment‑design suggestions. We conclude with a brief outlook on a causal agent that orchestrates data ingestion, hypothesis formation, intervention‑effect estimation, shift monitoring, and policy revision, offering an integrated, yet auditable and modular, workflow for reliability‑centered decision support.