How can we use AI to reduce errors (hallucinations)?