As I understand it, ContentID is notoriously over-aggressive in matching and only "works" in the sense that YouTube's interest in having it isn't particularly harmed by that, since its mostly a tool to improve relations and avoid lawsuits from big media interests.

Without manual validation, it probably wouldn't be a good model for identifying and removing ads from podcasts, especially using a "repeated in multiple podcasts" model, which doesn't start with known ads.