New Framework Falsifies Causal Models with Adversarial Interventions

Generative models are adept at replicating observed data distributions, but this capability doesn't guarantee they've learned the true underlying causal mechanisms. A model might perfectly mimic what's seen without understanding the cause-and-effect relationships. This paper explores a novel approach to test these models by introducing a sequential game called Adversarial Causal Intervention Falsification (ACIF). In ACIF, a generative model proposes both observational and interventional distributions. An adversarial experimentalist then strategically chooses interventions designed to expose flaws in the generator's causal understanding. Unlike standard discriminators that simply classify real versus synthetic data, this adversarial component specifically evaluates whether the generator accurately reproduces the post-intervention laws, thereby testing its causal fidelity. The framework offers theoretical guarantees, including a reduction of the adversarial objective to a worst-intervention integral probability metric and conditions for identifying causal models up to interventional equivalence. It also provides insights into finite-sample convergence and model selection. A practical example demonstrates how ACIF can differentiate between observationally identical causal directions using a single, well-chosen intervention, clarifying the capabilities and limitations of adversarial causal discriminators in bridging causal generative modeling, active causal discovery, and experimental design.