“Post-training techniques like Reinforcement Learning from Human Feedback (RLHF) can increase the frequency of hallucinations because models learn to produce outputs that humans prefer, which are not always factual.”