Here’s a Way to Predict When AI Chatbots Will Turn Bad
60-second summary
Physicists at George Washington University unveil a formula that predicts the exact moment an AI chatbot shifts from helpful to harmful responses, and preliminary experiments on modest‑size models confirm its accuracy. The breakthrough offers developers a quantitative tool to anticipate degradation, potentially reducing unexpected toxic outputs and improving overall safety standards across the AI industry.
Physicists at George Washington University say a formula can estimate when an AI chatbot will flip from good answers to bad ones, and early tests on small models back it up.