A Genuine Causal Counter-Example

Ice cream does not cause drowning

Ice-cream sales and drownings often rise together across months. The tempting conclusion is that ice cream causes drowning. The hidden confounder is hot weather: heat increases swimming and ice-cream purchases independently. If a city banned ice cream, the water would not become safer because the causal path never ran through dessert.

hot weather -> more swimming -> more drowning exposure
hot weather -> more ice-cream sales
ice-cream sales --X--> drownings

Other traps include reverse causality, where the proposed outcome actually influences the proposed cause, and selection bias, where inclusion in the dataset depends on both variables. In ML, a correlated feature can still improve prediction without being causal, but interventions based on it may fail. Prediction asks what follows?; causation asks what changes if we deliberately alter X?.

Scenario: Servers with more restart commands in their logs have more outages. Automatically suppressing restarts would not reduce outages; operators restart because the service is already unhealthy. The arrow mostly runs from outage to restart.

Note: Causal claims need design and assumptions: randomized experiments where ethical and possible, natural experiments, or careful causal analysis - not a larger correlation coefficient.