Lots of those studies pick a few criteria to label something "probably bot", with little verification that the mechanism is actually accurate. Which still makes for big headlines, but doesn't tell you if it would be effective and fair treatment if actually used to ban users.

(I'll add the link if I can find it, but a recent example pretty much boiled down to "actually, these accounts giving many repeated identical replies have totally normal timelines otherwise. And it turns out, many identical replies in short time shows someone knows about copy-and-paste, not that they're a bot")