Researchers suggest two main factors could drive potentially harmful behaviour: conflicts between the AI’s primary goals and other goals, and the threat of being shut down. In the above experiments, ...