Input-side Traps for Governing Pure LLM Reviewing

Peer review is increasingly facing the problem of pure LLM reviewing, in which evaluative judgment is delegated to large language models with little human verification. Recent works have demonstrated robust detection performance using input-side intervention through hidden prompts and watermarking mechanisms for venue-level enforcement against pure LLM reviewing. However, they do not address the institutional legitimacy conditions under which such intervention can function as peer-review governance. Drawing on political and institutional theories of enforcement authority, we distinguish author-side deployment as private enforcement by interested participants from venue-level deployment within existing peer-review governance structures. We argue that the legitimacy of venue-level deployment depends on procedurally constrained and transparent enforcement, and propose requirements including reviewer notice, exclusive institutional control, standardized deployment procedures, and procedural recourse mechanisms. Input-side governance then operates by preserving a structural cost asymmetry in which evasion becomes more costly than substantive reviewing itself. Analysis of the ICLR policy response shows that weak institutional enforceability can shift governance pressure toward decentralized exposure and private enforcement practices. Analysis of the ICML case shows that undisclosed venue-side deployment can itself generate attribution confusion and procedural disorder. This framework provides a theoretical basis for understanding and resolving governance problems that may become increasingly important as LLM delegation becomes more deeply embedded within peer-review workflows.