Formula predicts when AI chatbots are at risk of turning bad

AI chatbots: AI chatbots are conversational artificial-intelligence systems that generate responses based on user prompts and the context of an ongoing interaction. The reported research examines how their outputs can shift from acceptable to potentially dangerous responses and proposes a mathematical formula for identifying that transition in advance. Research finding: A physics-based model linked chatbot failures to instability in an attention mechanism and predicted the transition from safe to undesirable responses across publicly available AI models. Safety implication: The proposed method could support monitoring and testing systems by warning when a chatbot is approaching a potentially harmful response, including misinformation or unsafe guidance. Conversation context: Reordering the same questions changed whether models produced acceptable or undesirable answers, indicating that conversation history can influence safety outcomes.