Often times the words produced do have legitimate factual information though. It's less psychosis and more a confluence of well known human tendencies - salience bias, automation bias, etc.

The big problem is often times they don't as well. That's why we can't rely on them.