Correlation vs Causation

Correlation is a statistical relationship: as one variable changes, another tends to change too, measured for instance by a correlation coefficient. Causation means a change in one variable actually brings about the change in the other. The core reason correlation does not imply causation is that an observed statistical association can arise through several non-causal paths. A confounding variable can independently influence both measured variables, creating an association between them even though neither directly causes the other, as with ice cream sales and drownings both driven by summer heat. Reverse causation is a second path: assuming variable A causes B when in fact B causes A, such as assuming police presence causes crime rather than crime causing more hiring. A third path is coincidence or spurious correlation, where two unrelated trends move together purely by chance, especially likely when many variables are tested against each other, a problem statisticians call data dredging. This distinction is central to how machine learning is used responsibly. Predictive models, including large-scale AI systems trained on observational data, are fundamentally correlation detectors: they learn statistical patterns in their training data without inherently understanding cause and effect. When such patterns are used to justify interventions, such as denying loans, targeting policing, or setting insurance prices based on correlated but non-causal factors, the result can encode and amplify existing biases in the data rather than identifying true causal drivers. Establishing genuine causation typically requires controlled experiments or specific statistical techniques designed to account for confounding, rather than pattern-matching on observational data alone.