Stick to your Role! Stability of Personal Values Expressed in Large Language Models