Causation instead of correlation

In short: Two things can occur together without one causing the other. Machine learning finds correlations. Whether an intervention achieves something is a question of causation.

Classic traps

• Common cause (confounder): ice cream sales and sunburns rise together, both depend on sunny weather. Ice cream does not cause sunburn.
• Reverse direction: people who take a lot of medicine are ill more often. The illness is the cause, not the medicine.
• Chance: with very many comparisons random links always turn up.

How to check causation

1. Randomised experiments (A/B tests): you divide randomly into groups. Because chance decides, confounders even out on average. This is the gold standard.
2. Observational data with assumptions: you model cause-effect relations, often as a causal graph, and adjust for confounders (regression, matching, instrumental variables).
3. Natural experiments: random differences in reality, such as a rule change in only one region.

Why this matters for AI

A model that only knows correlations breaks when the world changes. And whoever reads explanations from SHAP and LIME as causes draws wrong conclusions. For bias and fairness too the question of causes is central, see Bias, fairness and data quality.

See also: Probability, entropy and KL divergence