Under the classical epidemiological definition, a confounder of $X$ and $Y$ is any variable $Z$ that is associated with $X$ and (among the unexposed) with $Y$, and that does not lie on the causal path from $X$ to $Y$. M-bias shows these association-based criteria are insufficient. In the structure A leftarrow B rightarrow C leftarrow D rightarrow E (with exposure $A$ and outcome $E$), the middle variable $C$ is statistically associated with both the exposure and the outcome and does not lie on any causal path between them, so the classical definition would label $C$ a confounder. Yet $C$ is a collider: adjusting for it opens a spurious path and introduces bias instead of removing it. Because the classical criteria cannot distinguish such colliders from genuine confounders, they fail to capture the causal structure needed to define confounding correctly.