Do GenAI avatars open new responsibility gaps?

Abstract In this article, I argue that semi-autonomous avatars relying on generative artificial intelligence to replicate or represent real human persons—GenAI avatars—open a new type of responsibility gaps, which I call “proxy gaps”. Proxy gaps refer to situations when we cannot hold anyone morally responsible for the outcomes of GenAI avatars, because the representation relationship between avatars and humans is shaped by multimodal Large Language Models (LLMs). In addition to epistemic gaps by AI avatars discussed in the literature, I argue that GenAI avatars also open control gaps—where no one really controls the output of the avatar. I introduce the “proxy-control paradox” to explain why control gaps arise: in trying to achieve improved control over the desired outcome of their avatar, humans need to delegate control over the process leading to that outcome to the GenAI technology. Together, the epistemic and control gaps complicate the two criteria traditionally used for moral responsibility, resulting in a proxy gap by GenAI avatars. Despite inherent proxy gaps, I argue that, under certain circumstances, we can still rightfully hold individuals morally responsible for the outcome of their GenAI avatars. I detail four conditions pertaining to human understanding and LLM personalization, as well as the right to veto and outcome control, which, taken together, can ground individual moral responsibility for the outcome of personal GenAI avatars.