Academic journal publishing is undergoing a significant transformation driven by artificial intelligence, which is being applied across all stages—including authors' writing, editors' evaluation, peer review, and comprehensive screening by academic publishing platforms. Technological empowerment is a double-edged sword: it brings both substantial benefits and considerable challenges. This study focuses on AI-assisted peer review and examines the following liability questions: who is responsible when AI delivers erroneous judgments, discriminates against specific research fields or author groups due to biases in its training data, accesses copyright-protected academic literature without authorization during training and operation, or causes the leakage of authors' unpublished research data and personal information—is it the AI developer or the journal that deploys the system? This study sorts AI-assisted review into three types of use cases, each with its own technical features and typical scenarios. The study finds that legal risks fall into four areas: copyright infringement, reputation rights infringement, culpa in contrahendo and breach of contract, and personal data security. When AI reconfigures the legal relationship among authors, journals, and reviewers from a triangular into a diamond-shaped structure, three key dilemmas emerge: responsible-party ambiguity and the absence of fault-determination standards; complex causality and an unbalanced burden of proof; and heightened difficulty of liability attribution due to algorithmic bias. Therefore, to solve these issues, this study proposes a differentiated liability framework to address these distinct risk categories, clarifying attribution principles and liability mechanisms for journals, AI developers, and other stakeholders. For copyright infringement risks at different stages, the core criterion is to determine responsibility based on the controlling power of the relevant subject over the infringing act. For reputation rights infringement caused by AI, the presumption of fault principle should be adopted, with the journal as the primary liable party, while also taking the fault-based liability of the AI developer into account. For culpa in contrahendo and breach of contract, the principle of privity of contract should be strictly followed, with the journal as the core liable party, and the constitutive elements and liability methods should be specified according to the type of liability. On the institutional side, a package of reforms across legislation, judicial practice, and industry self-regulation, including a classified filing system, reversal of the burden of proof, and establishment of ethics committees—is proposed.