Decentralized Physical AI systems—where autonomous agents interact with the physical world while relying on decentralized coordination layers—create legal frictions that traditional AI governance frameworks do not fully capture. This talk isolates two structural challenges. First, the lack of legal personality: in fully decentralized settings, it may be difficult to identify a legally recognized counterpart able to enter binding contracts, hold assets, or be sued. Second, the liability gap: as autonomy and decentralization increase, attribution of harm becomes harder, pushing legal systems toward “scapegoat” solutions (human-in-the-loop or proxy actors) not because they are operationally necessary, but because they are liable.

Building on EU liability discussions and the limits of “human-centric” approaches, the session proposes reframing decentralized AI as infrastructure rather than a firm-like actor. Under this perspective, risk management shifts from searching for an elusive developer/owner to a layered model combining ex-ante protocol authorization, mandatory insurance mechanisms, and public authorization labels to support trust and market discipline. The outcome is a concrete, regulatory blueprint aimed at reducing uncertainty without freezing innovation.