Scholarly Communication as an Engineering Discipline

Abstract

Scholarly communication has long relied on social filters: when writing and careful presentation were costly, polished prose and venue prestige served as imperfect signals of effort and quality. Large language models (LLMs) erode these signals by making fluent text cheap and enabling rapid, large-scale production of plausible manuscripts, including adversarial mimics designed for surface plausibility. Meanwhile, attention and verification remain scarce. We argue that scholarly communication has entered a communication-constrained regime. We formalize publishing as distributed inference with many encoders (studies) and many decoders (readers) operating under attention budgets and adversarial contamination. The model motivates a separation of representation from responsibility: machines may render prose, but authors remain accountable for structured claims, evidence, and uncertainty estimates. We derive limits showing why binary gatekeeping and implicit routing fail under overload and shifting base rates, and we translate them into requirements for the Protocol for Post-LLM Science: claim-evidence records, graded (probabilistic) certification, verification-aware routing, and explicit congestion control for scarce verification capacity.