Researchers at the University of Vermont have discovered that human language exhibits a deep safety bias, challenging a 70-year-old scientific consensus in psychology, linguistics, and AI.