When Scale Matters: Quantifying Causal Distortion in Multi-level Behavioural Systems

The set-wise, multi-level nature of human behaviour presents a fundamental barrier for Artificial Intelligence (AI), as prevailing models lack a first-principles theory to capture its true causal architecture. This absence forces a reliance on pairwise approximations and heuristic tuning, leading to biased and unreliable inferences. To overcome this, we introduce the algebraic theory of causal granularity—a foundational framework that formally describes how causality operates and transforms across behavioural scales. From this basis, we derive three principled metrics: causal complexity, granularity gap, and collective causal entropy, which directly quantify multi-factor interactions, information loss from aggregation, and system-level uncertainty. Our work thus provides the missing theoretical foundation for AI to achieve true alignment with human behaviour, enabling robust and interpretable causal inference. This represents a pivotal advance, moving behavioural AI from correlation-driven approximation to a first-principles science with direct implications for addressing complex societal challenges.