Causal Articulation Theory (CAT): Articulating Static and Temporal Causal Models Beyond LLM Rationalizations

Matej Zečević, Devendra Singh Dhami, Kristian Kersting

Keywords: causal explanation, structural causal model, interactive machine learning, large language model

Abstract: In normative terms, a meaningful explanation should reflect how data is generated—this is precisely where causality becomes essential. Existing machine learning methods for explanation either miss this desideratum entirely or incorporate only partial causal knowledge. Large language models (LLMs), in particular, routinely produce fluent explanations that amount to unfaithful rationalizations of the underlying data-generating process. We therefore advocate a direct approach: derive explanations from the structural parameters and mechanisms of a causal model itself. To this end, we introduce Causal Articulation Theory (CAT), a formal account of how explanations can be articulated from a structural causal model (SCM). CAT addresses why-questions about individual units and uses a recursive articulation procedure that draws on both the graphical structure and causal effects encoded in the SCM. Using a small set of first-order articulation rules, we first develop CAT for static linear SCMs and show that it offers an appealing alternative to LLM rationalizations: CAT naturally distinguishes direct from indirect causes, captures the qualitative sign of causal effects, and remains robust to stochasticity in exogenous variables. To address ever-changing environments, we then extend CAT beyond static settings by relaxing initial assumptions to cover both temporal and agentic scenarios, yielding articulated explanations over time and action. For empirical corroboration, we present a series of experiments: (i) a user study examining the alignment between CAT-based articulations and human causal judgments in everyday domains, (ii) an investigation of CAT as a regularizer for causal discovery, and (iii) examples of articulated explanations in two temporal domains involving forecasting and a simple video game environment.

Submission Number: 64