Does AI create a coupled arms race in research production and review?

The survey argues that AI's effect on research production and on peer review should be read as one coupled process, not two separate trends. It synthesizes 230 scholarly publications and institutional records through a taxonomy of six linked dynamics: production scaling, evaluation automation, evaluation manipulation, defense mechanisms and policy responses, evasion and side effects, and long-horizon ecosystem feedback. Its summary of the progression is that "cheaper and faster research production increases pressure on evaluation, AI-mediated evaluation becomes more scalable and repeatable, participants can exploit evaluator regularities, and institutions respond with technical safeguards and policy controls." The survey is explicit that its confidence varies. Evidence is "strongest for production and evaluation at scale, reproducible manipulation, and institutional response," while post-policy adaptation and long-horizon feedback "remain less directly observed."

The mechanism the survey gives is relational. Production scaling changes the demand placed on evaluation. Automation makes parts of evaluation "increasingly repeatable and therefore more susceptible to manipulation and optimization." Venues respond with defenses, and those defenses "can induce evasion or shift costs and risks to other actors." The taxonomy is organized around response relations among actors rather than around technologies. An arms-race episode is defined narrowly: an action that changes a signal another actor relies on, a counter-response, and a resulting shift in incentives. Because complete sequences are rarely observed in one study, the survey says connections drawn across separate literatures "are presented as research questions rather than as directly observed causal sequences." Its discussion draws the practical conclusion that trustworthy evaluation is the bottleneck. "Faster review generation is useful only when the resulting judgments remain grounded."

Against the nearby notes, the survey shares a premise with Can human review keep pace with AI-accelerated research generation?: once generation gets cheap, review has to scale too. Where that note offers an ordered ladder of collaboration levels, this survey offers six dynamics linked by response relations, so the same pressure shows up as a feedback structure rather than a sequence of roles. The evaluation-manipulation dynamic has a concrete counterpart in How much does rhetorical style shift AI review scores?. That result shows scores moving with presentation while content is held fixed, which is the kind of change the survey's section on adaptation says makes static evaluation unreliable. On the defense side, Does banning LLM use in peer review change review outcomes? reports substantial noncompliance at one venue. That is the evasion the survey's defense discussion anticipates, though the survey does not measure it.

What the excerpt does not establish is as important as what it claims. It is a synthesis of existing literature and adds no new measurements of its own. The survey also says its framing "is intended as a lens for organizing these interactions rather than as a claim that all AI-assisted research, reviewing, revision, or institutional change is adversarial." The excerpt's conclusion heading is empty, so nothing here rests on a closing summary. The implication is modest. The six-dynamic map is useful for asking which response relation a given finding belongs to, and "arms race" should be treated as a hypothesis about sequence until someone follows the same venue over time, which the survey itself says is the thinnest part of the evidence.