The quantified likelihood that an AI system will make a harmful or incorrect decision in real-world deployment.