Ensuring the Alignment of Genetic/Epigenetic Designed Swarms

One of the major concerns of AI researchers and implementers is how to ensure that the systems stay aligned with the aspirations of the humans they interact with. This problem becomes even more complex for systems that develop their own operational rules and where multiple agents are involved. The paper addresses some of the implications of using genetic/epigenetic design techniques where the control structure is developed without direct human involvement. This presents particular difficulties in ensuring that the control protocols stay aligned with the desires of the instigators and do not cause unpredicted harm. It also explores how this problem is further complicated when the AI system has many agents. Modern control systems are often decentralized which provides a more robust solution than using a central controller. A specific example of this approach is Self-Organising Swarms where the agents act independently of the central control. From an alignment point of view, it generates particular problems. Not only must the individual agents act in the best human interest but the swarm as a collective must do it as well. This is difficult for a homogeneous swarm and no proposal for a heterogeneous one has yet been made. There have been and continue to be considerable research and discussions on how to create and what form a global AI ethics might take, but any progress has been slow. This is partly because even the Universal Declaration of Human Rights has difficulties. All the nations that have signed up to the UN Human Rights Declaration believe they are at least trying to implement it. The problem is in the interpretation where many signatories believe others are in breach. The same would apply to any universal AI ethics agreement. This paper proposes a solution where the AI systems’ basic ethics are individual but have to comply where they interface with either other AI entities or humans.