Bottom line: Asks what it would mean for an AI to be truly accountable. Argues that accountability needs a persistent self where consequences can land and be remembered — otherwise the burden falls on nearby humans.
Asks what it would mean for an AI agent to be accountable as a continuing entity. Introduces "consequence reception" as a design requirement: accountability needs a persistent boundary where consequences can land, be remembered, and shape future action. Without that, the burden of correction and blame defaults to nearby humans. The paper argues that present LLM agents lack four properties needed for genuine accountability: boundary, accumulation, consolidation, and action-guiding substrate.
If a robot makes a mistake and we just reset it, who learned the lesson? The robot didn't — it's a blank slate. The human who has to deal with the consequences did. This paper says that for an AI to be truly accountable, it needs to have a persistent self where consequences can actually land — where it can feel the weight of its mistakes and grow from them. Without that, "accountability" is just a word that shifts the burden onto humans.