Flip a light switch and the bulb turns on; flip it back and the bulb turns off. You changed one thing, the switch, and the bulb changed every time. That is the test for causation: if you deliberately change X while keeping everything else the same, does Y change? Watching two variables move together is a weaker test, because you did not do the changing.
In observational data, three things can create a correlation without X causing Y. A confounder, or lurking variable, can drive both: hot weather raises ice cream sales and also sends more people swimming, so ice cream sales correlate with drownings. The direction can be reversed: a city with more crime hires more police, so police numbers and crime rise together even if police reduce crime. And among many variables, some pairs line up by coincidence. Each alternative has to be ruled out before you claim a cause.
Randomized experiments handle all three at once. Split 200 volunteers into two groups by coin flip, give one group the treatment and the other a placebo, and the coin flip makes the groups alike on average in everything else, including traits nobody thought to measure. If the groups then differ by more than chance allows, the treatment is the best explanation. Compare dependence: knowing A can shift the odds of B because of a confounder, while causation is the stronger claim that intervening on X moves Y.
Think of it like this. Causation is the difference between a thermostat and a thermometer. A room's thermometer reading and its temperature rise and fall together, yet warming the thermometer with your hand changes only the reading. Turning up the thermostat actually warms the room. A correlation may be a thermometer, merely tracking; a cause is a lever. In real data, a cause usually shifts Y on average, not every time.
Core idea. Causation means altering one variable directly produces a change in another, beyond just moving together.