Counterfactual Reasoning as the Missing Link Between Statistical AI and Human Cognition: A Cognitive Science Perspective

Text-only large language models (LLMs) used as general-purpose reasoners achieve remarkable performance on linguistic and reasoning tasks yet remain limited on intervention-sensitive and individual-level counterfactual reasoning. This article offers a cross-level synthesis linking formal causal competence, cognitive simulation, neural implementation, and artificial intelligence (AI) design criteria. Drawing on formal causal inference theory, developmental cognitive science, and cognitive neuroscience, we first characterize the three levels of causal cognition—association, intervention, and counterfactual reasoning—and show why each is formally irreducible to the one below it. We then discuss concrete, domain-specific categories of LLM failure that are consistent with the absence of a structured causal model capable of supporting individual-level counterfactual invariance. We review neuroscientific evidence for component processes relevant to counterfactual cognition—including episodic construction, fictive evaluation, and internally generated scenario simulation—with the Default Mode Network and hippocampal systems playing important but not exclusive roles. Finally, we propose that the most productive path forward is not to implement Pearl’s formal rules in AI systems, but to study how the brain approximates counterfactual reasoning without ever following those rules explicitly—just as neuroscience has historically provided useful constraints, architectural motifs, and evaluation criteria for AI research.