Language models predict the next word in a sequence.

Training data teaches them which words usually appear together.

That's basically it in a nutshell.

Truth and accuracy? Those are just happy accidents when the patterns happen to be correct.