It’s because when you take small sample size of “people on the internet” you’ll get either completely random set of beliefs and assumptions, or you’ll get beliefs and assumptions that you’re prejudiced to project onto the wider sample.