Who bears responsibility when AI systems cause harm?

This paper identifies a fundamental gap in AI accountability: harm occurs, systems are identified, but no continuing agent experiences consequences that would change future behavior. The authors argue that punishment theories—deterrence, rehabilitation, retribution, incapacitation—all require a persistent entity that registers corrective signals and updates accordingly. Large language model agents, as software composites freely copied and reset, cannot satisfy this requirement. Current legal frameworks fail: either humans bear accountability for behaviors outside their control (the 'moral crumple zone'), or AI entities are created without ensuring the decision architecture receives feedback as a behavioral signal. The paper positions consequence-agency coupling as a sociotechnical infrastructure problem, concluding that until such architectures exist, high-stakes AI systems should remain under human principals with meaningful control and liability.