In this episode of our series about Metaphysics and modern AI, we break causality down to first principles and explain how to tell factual mechanisms from convincing correlations.
From gold-standard Randomized Control Trials (RCT) to natural experiments and counterfactuals, we map the tools that build trustworthy models and safer AI.
In the fourth episode of our metaphysics series, we explore one of science's most fundamental questions: what are causal relationships?
Causality, or etiology—the study of causes—requires three key conditions for any cause A and effect B:
Causal relationships take different forms, from simple chains (A to B to C) to homeostatic loops, common effects, and common causes
Our minds are wired to perceive patterns as meaningful connections—a principle explained by Gestalt theory, which suggests we perceive unified wholes rather than individual parts. This leads to spurious correlations, like the 98.5% correlation between arcade revenue and computer science PhDs awarded that's discussed in this episode.
.
As Judea Pearl emphasized in Causality: Models, Reasoning, and Inference, correlation does not imply causation. Statistical associations tell us measures trend together but reveal nothing about how they're linked.
The distinction is critical:
The gold standard is the Randomized Control Trial (RCT), where participants receive randomized treatments under controlled conditions. When RCTs aren't possible, researchers turn to natural experiments or quasi-experimental designs, which retrospectively analyze data to tease out causal relations.
Studying causation remains challenging because it's a form of inference where effects might have alternative explanations, and findings may result from chance without proper statistical controls. David Hume raised the skeptical argument that all causality is inferred—we can't truly disambiguate a series of actions from a necessary chain of events.
As we build AI models, we must remain vigilant about what we're actually modeling. Are we modeling correlational findings? Then predictive models work well. Are we modeling cause and effect? Then we're building a physical model. Causality is the glue that ties random observations into a coherent understanding of the mechanistic parts of our world, which is essential knowledge for anyone working in AI and data science.
This is the fourth episode in our metaphysics series. Each topic in the series is leading to the fundamental question, "Should AI try to think?"
Check out previous episodes:
If conversations like this sharpen your curiosity and help you think more clearly about complex systems, then step away from your keyboard and enjoy this journey with us.
