Next token prediction is a pretraining objective that doesn't tell anything about behavior and activation structure of the resultant network. The literature that explores a hallucination problem has little to do with your claim about physical impossibility.