With the rapid development of autonomous decision-making systems, generative artificial intelligence, embodied robots, and algorithmic agents, criminal law is increasingly confronted with the question of whether artificial intelligence can or should be recognized as a criminal subject. Traditional criminal law is built upon human-centered assumptions: a criminal subject must possess legal capacity, behavioral capacity, culpability, and the ability to bear punishment. However, advanced AI systems may independently generate harmful outputs, manipulate information, control physical devices, or participate in decision-making processes that lead to legally significant harm. This paper adopts a systematic review method to examine the criminal subject qualification of artificial intelligence from the perspectives of legal personhood, actus reus, mens rea, culpability, punishment, and accountability allocation. It argues that current AI systems should not be recognized as independent criminal subjects in the full sense because they lack consciousness, moral self-reflection, normative understanding, and punishment responsiveness. Nevertheless, AI challenges the traditional binary structure of “human subject–technical object” and requires a reconstructed accountability framework. The paper proposes a layered model of criminal responsibility in which developers, deployers, operators, corporate entities, and AI systems are functionally distinguished. AI should be treated as a risk-generating autonomous instrument rather than a punishable offender, while criminal law should strengthen duties of design, supervision, risk assessment, explainability, and post-incident traceability. This approach can avoid both excessive anthropomorphism and responsibility gaps, thereby preserving the normative foundation of criminal law while responding to the practical risks created by artificial intelligence.