They just believe right now.  They are just parroting whatever is given to them. 

Many models done by big corporations go thru two main phases:
1. pre training (everything on the internet except maybe truth from deplatformed people)
2. supervised fine tuning (adding skills to models and also injecting lies).
Many models done by big corps have lots of truth and lots of lies. The trivial truth and important lies.

I measure the misinformation and I have a leaderboard about this.  I can DM you some research if you want.

And I also build a truth seeking model: Ostrich. 

LLMs don't have the ability to differentiate between facts and opinions. I believe LLMs should still learn from humans because humans have souls which is connected to God which insert the feeling that something is FACT and not an opinion.