ABSTRACT

We distinguish artificial intelligence (AI) as another automation technology from AI as a new creature posing existential threats. Then, we abstract the existential threats of AI by examining their essential nature, and argue that (i) these threats are best understood as endemic to all intelligences, and (ii) they can be addressed using solutions derived from human (political and cultural) history as well as the evolutionary biological history. We discuss how many of the threats are fundamentally due to a power imbalance between AI and humans. This enables us to transform a large, ambiguous, part of the AI security problem to a familiar political problem since politics is primarily about power. This transformation enables us to address many AI threats by reusing some of the structures and procedures humans developed over time to manage power in human societies. This and other insights lead to proposing principles that must be embedded in AI and imposed on AI developers and operators. The principles belong to four categories: building safe power structures, maintaining operational stability and control, enhancing our security through engineered AI psychology and procedural principles that govern the AI security practice itself. For instance, we argue that all AI entities must be discrete, finite, mortal, supervised, etc., and that their power should be distributed and subject to governance. Moreover, we insist on the general philosophical attitude that all intelligence is inherently unpredictable, dangerous if unchecked, and that the more an AI entity is capable in the physical world, the more it must adhere to human-like and genetic-like limitations and standards of behavior. And finally, we strongly advocate building an AI security and safety practice that involves experts from the humanities and biology as well as technical experts in AI and computing.