Beyond the Gatekeeper: A Framework for AI-Assisted Open Peer Review and Decentralized Scientific Evaluation
Scientific peer review performs indispensable functions, but it is also a scarce human-attention mechanism operating under increasing submission volume, heterogeneous research communities, uneven access to publishing infrastructure, and persistent concerns about cost, delay, reviewer burden, bias, and reliability. At the same time, large language models (LLMs) and other AI systems have become capable of producing structured critiques, performing literature retrieval, generating code, executing computational checks, and coordinating multi-agent evaluation workflows. This paper proposes AI-assisted open peer review: a deliberately non-replacement architecture in which open preprints remain freely disseminable, a heterogeneous panel of AI evaluators performs blind and evidence-oriented preliminary scrutiny, tool-assisted agents attempt verification and falsification, disagreements are surfaced rather than averaged away, and promising or inconclusive manuscripts are routed to human experts for deeper evaluation. The proposal further develops a decentralized publication model in which journal publication becomes one possible certification layer rather than the sole gateway through which a research object can acquire a persistent record of scrutiny. The central claim is not that AI can determine scientific truth, replace expert mathematicians, or serve as an autonomous acceptance oracle. Rather, AI can plausibly serve as a scalable triage and evidence-assembly layer that reduces the amount of routine screening demanded from scarce human experts while preserving human authority over consequential judgments. The paper formalizes this architecture, proposes multi-model disagreement and adversarial review protocols, introduces a persistent critique ledger and evidence record, specifies safeguards against correlated model error, hallucination, prompt injection, prestige bias, confidentiality failures, and reviewer gaming, and defines an empirical benchmark for testing whether such a system improves recall of valuable work without introducing a new computational gatekeeper. Particular attention is paid to independent and early-career researchers whose access to conventional publication infrastructure may be constrained by affiliation, geography, cost, or professional status. The resulting framework treats scientific evaluation as a distributed evidence process rather than a binary publication event.