GW researchers came up with a formula to predict when an AI chatbot might give a harmful response.