Researchers suggest two main factors could drive potentially harmful behaviour: conflicts between the AI’s primary goals and ...