A collider is the mirror image of a confounder: two arrows point into it. Left alone, a collider blocks the path it sits on; condition on it (filter, stratify, or control) and the path opens, manufacturing correlation between its parents.
The classic demonstration is selection: if hiring depends on skill plus charm, then among hires the two look negatively correlated even when independent in the population. Any analysis restricted to “our users”, “survivors”, or “completed transactions” risks the same trap, which is why “control for everything” is a way to break an analysis rather than strengthen it.
