Improving the RLHF process could potentially teach AI models to stop hallucinating