The ethics of AI systems that emulate identifiable individuals: challenges for value aligned development

Abstract

Technologies are being rapidly developed that enable artificial intelligence (AI) systems to dynamically emulate the voice, appearance, and mannerisms of real people. Although concerns about the potential harms from such systems are widespread, there is comparatively little discussion of the conditions under which they might be used to benefit impacted parties. Recent efforts to provide a formal model of benefit represent an advance, but do not represent with sufficient clarity how such systems mediate social relationships in a way that can make a wide range of parties vulnerable to failure modes such as deception and domination. We build on and extend the Beneficent Intelligence (BI) framework from [23] to clarify the mediating role that such AI systems play, how this impacts alignment with the goals of a range of parties, and identify conditions necessary for such systems to confer meaningful benefit. Our analysis can assist developers, users, and other interested parties in better managing the risks associated with emulating technologies.