The consequences of super intelligent AI are unpredictable, and the vast majority of randomly specified utility functions do not have optima with humans in them, making it difficult to ensure that AI remains aligned with human values.