Journals are experimenting with AI-assisted peer review, but reviewers could game it and authors could optimize for the model instead of the science. Does AI review improve rigor, or does it just add a new attack surface to an already strained system?