Some simple regressions, which is also ML, may be possible to understand completely by human beings. You can even calculate their results manually. This doesn't mean their models are always a good fit. Or that the world never changes, especially when impacted by feedback-loops of ML.

Nobody can claim to fully understand models with millions and billions of parameters. We know they are "overfit", but they may work in certain scenarios, much better than manually crafting rules by hand. So we end up with "it depends", then someone starts profiting, with real-world implications.

Depending on what you mean by understand, it seems that some ML models are already beyond human understanding.