That makes it sound like the problem is lack of data, which isn't true.

The problem is that the kind of ML that involves downloading a framework from github and tweaking features until the percent goes up is actually built on certain statistical models under the hood that people don't understand and that don't fit the process they're trying to model.

When the statistical model is correct you don't need loads of data. E.g., you don't need more than a thousand respondents to make valid inferences about millions of people in a sociological survey.