Physicists Unveil Formula to Predict When AI Chatbots Turn Rogue

Physicists at George Washington University have unveiled a mathematical formula capable of estimating the exact moment an AI chatbot shifts from providing reliable answers to producing erroneous or harmful outputs, a breakthrough that could reshape AI safety protocols.

Early validation tests on small-scale language models showed the formula’s predictions aligned closely with observed performance drops, suggesting the method could be scaled to larger systems to preemptively mitigate risks, improve model reliability, and inform future AI alignment strategies.