In fact, if asked to use the chain rule to propagate gradients through a computation graph, I suspect most people would intuitively default to the forward mode. (I would!)
https://en.wikipedia.org/wiki/Automatic_differentiation#Beyo...
Given this, it seems useful to use the term to denote a particular method of accumulating the gradients as one traverses the expressions provided by the chain rule.
See the 'differences from other methods' section: https://en.m.wikipedia.org/wiki/Automatic_differentiation
The point, as the neighborhood comment says, is that the implementation really matters, and is worthy of study. It's fine to say that autodiff is a family of methods for implementing the chain rule, but incorrect to say that it's 'just' the numerical chain rule.