“Over the course of AI development, safety properties like controllability and theoretical justification have been sacrificed for higher performance.”