Peer review relies on independent scientific judgment, yet the increasing use of generative artificial intelligence (AI) raises concerns about accountability, confidentiality, and the appropriate division of responsibilities between human reviewers and AI systems. This Perspective proposes a function-based framework that distinguishes three forms of AI involvement: AI-assisted language editing, AI-supported verification, and AI-generated scientific judgment. Language editing may be generally acceptable when reviewers retain ownership of their reasoning, while verification tasks require independent assessment of AI-generated findings. In contrast, relying on AI to formulate substantive criticism or publication recommendations without independent evaluation risks epistemic delegation, whereby professional authority becomes detached from defensible scientific judgment. The proposed framework emphasizes that the ethical boundary should depend on the function assigned to AI rather than the mere presence or extent of its use. It also offers practical recommendations for journal policies concerning confidentiality, verification, disclosure, and reviewer accountability. Reviewers should be able to trace AI-assisted criticisms to the manuscript, verify their validity, explain their reasoning, and defend their conclusions. By distinguishing acceptable assistance from inappropriate delegation, this framework provides a practical basis for integrating AI into peer review while preserving independent human judgment and the integrity of scientific evaluation.