哆啦A夢是民法上的物嗎？──從大雄偷窺事件談起

Doraemon as a Thing in Civil Code?: A Discussion Starting from Nobita’s Peeping Incident

本研究以民法解釋論提出涵蓋自然人、法人、AI系統或羸機器人、強AI系統或機器人之侵權行為四元責任主體系譜，並以有意義的人類控制（Meaningful Human Control）理論，充實民法上物定義的支配可能性。其次，本文主張支配可能性概念中的「可支配性」區分四大要件——支配主體存在、透明性、控制性、支配意志——以貫穿，調適傳統侵權行為法於AI時代的適用。針對AI系統／機器人，筆者強調（1）在設計階段應由人類明確賦權：透過程式碼法則與介面定義設下其行動範圍裁量框架應屬弱裁量模式；（2）在運行階段必須保有人類承責角色：設計者與使用者共同承擔系統失誤的法律責任；（3）援用諾齊克真理追蹤理論，要求系統在「P為真」與「P為假」兩種反事實情境下，對人類道德理由呈現一致回應，並由人類適時介入判斷真值與否；（4）確保人類對系統行為具備最終監督權限：包括即時介入、覆核、重置與糾錯。如此，雖AI擁有自動化決策能力，仍落在「人類意志控制」框架內，並透過理論上建構民法第184條侵權行為四元責任主體系譜、消費者保護法第7條、類推適用民法第188條等學理建構性詮釋，實現損害賠償保護被害人目的。此機制不僅維護法律可預見性，也兼顧被害人保護與技術創新。

This study proposes a four-tiered typology of tort liability subjects under Civil Code: natural persons, legal persons, AI systems or robots (weak AI), and strong AI systems or robots. Through a Civil Code interpretative approach, it aims to broaden the traditional definition of“thing”by incorporating the theory of Meaningful Human Control (MHC), with a particular focus on the notion of controllability. This notion is further articulated through four essential components: the existence of a controlling subject, transparency, actual controllability, and the intention to exert control. These elements collectively provide a normative structure for recalibrating tort liability doctrines in response to the evolving challenges of AI technology.
Focusing on AI systems or robots, this paper argues: (1) During the design phase, humans must be explicitly empowered by embedding rule-based frameworks and clearly defined interface boundaries, thereby constraining the AI’s operational discretion within a weak-discretion model; (2) During the operational phase, the assignment of legal responsibility must remain traceable to human agents—namely designers and users—who are jointly liable for failures arising from system behavior; (3) Inspired by Robert Nozick’s truth- tracking theory, AI systems should demonstrate consistent moral responsiveness across both actual and counterfactual scenarios, while humans retain the authority to determine the truth value of outcomes; (4) Human actors must preserve ultimate supervisory powers over AI behavior, including the capacity to intervene, review, reset, and correct errors in real time.
In this way, even though AI systems possess autonomous decision-making capabilities, their operation remains situated within a framework of Meaningful Human Control. This legal model is further substantiated through a constructive doctrinal interpretation of Civil Code Article 184, Article 7 of the Consumer Protection Act, and the analogical application of Civil Code Article 188. Ultimately, this framework enhances legal foreseeability and ensures that AI development progresses in tandem with the dual imperatives of protecting victims and supporting technological innovation.