One small adjustment to the method of machine learning allowed MIT researchers to largely fix AI exaggeration and hallucination.