AI code review, particularly using models like GPT-4, presents promising time-saving potential but faces significant challenges and limitations. Although AI can quickly generate reviews, helping developers receive immediate feedback and potentially reducing review cycles, its accuracy and reliability remain problematic. Internal experiments at Graphite revealed that AI reviewers often produce false positives and lack the contextual understanding necessary for effective code assessments. Efforts to improve AI performance involved customizing review guidelines and employing retrieval-augmented-generation to provide broader codebase context, which somewhat reduced errors but didn't adequately enhance the signal-to-noise ratio. More fundamentally, philosophical and practical issues, such as the need for human trust, accountability, and nuanced codebase context, suggest that AI lacks the subjective judgment required for comprehensive code review. While AI can complement human review processes by acting as a super-linter or providing contextual information, it is unlikely to replace human reviewers entirely in the foreseeable future, as the final approval of code changes will likely remain a human responsibility.