Ice cream sales and drowning deaths rise together every year. Ice cream doesn't drown anyone; summer causes both. Everyone laughs at this example and then walks into the same trap wearing a lab coat.
Correlation means two things move together. Causation means one produces the other. Data — any data, any amount — can only ever show you the first. The second requires knowing why, and why is not stored in the columns.
How machines fall in
Machine learning is, mechanically, the industrial-scale discovery of correlations. That's not a criticism; it's the job description. The trouble is that the model has no way to distinguish a deep pattern from an accidental one:
- A famous medical model learned that asthma patients had better pneumonia survival — because doctors, knowing the danger, rushed them into intensive care. The correlation was real. Acting on it as causation would kill people.
- An image classifier learned to spot huskies by detecting snow in the background. The correlation held in the training photos. It collapsed the moment a husky stood on grass.
- A hiring model learns that certain zip codes "predict" job success — laundering decades of housing discrimination into a clean-looking score.
These are called spurious correlations or shortcut learning, and more data often makes them worse, not better: bigger datasets contain more accidents to discover.
The question that protects you
When a model — or a chart, or a confident colleague — shows you a pattern, ask: what else could produce this? The drowning numbers don't need ice cream; they need summer. The husky detector doesn't need a husky; it needs snow. If you can tell a cheaper story that fits the same data, the pattern hasn't earned your trust yet.
What to read next
With the trap marked, you're ready for the core move of all machine learning: patterns and predictions — how finding regularities becomes the engine of everything.