A simple rule for causation

AI researcher Vivek Hebbar has published a significant paper titled 'A Simple Rule for Causation' on the LessWrong forum, introducing a practical heuristic for causal inference that could enhance AI reasoning capabilities. The central claim provides a testable condition: if every variable X found to be correlated with A is also correlated with B, we gain moderate confidence that A causes B. Conversely, finding even one X related to A but independent of B falsifies A→B causation. This addresses a fundamental challenge in AI—distinguishing causation from mere correlation using purely observational data, which is crucial for developing more robust world models.

The rule emerges from analyzing dependency structures in triples {A, B, X} and shows that while a 'common cause only' hypothesis can never be fully ruled out, the causal directions A→B and B→A can be falsified by a single contradictory observation. Hebbar details exceptions, including designed control systems like thermostats where causal paths may cancel out, creating statistical independence where causation exists. For AI systems, this heuristic offers a computationally feasible method to reason about causality without controlled experiments, potentially improving agent decision-making, scientific discovery pipelines, and the reliability of models that must understand intervention effects.