Preventing the Collapse of Peer Review Requires Verification-First AI

TL;DR

The paper addresses the risk that AI-enabled claim inflation could collapse peer review by shifting incentives toward proxy signals. It develops a minimal mixture model with latent truth $T$, observed score $S$, verification probability $q$, and proxy noise $\Delta$, deriving the truth-coupling $\rho=\mathrm{Corr}(S,T)=\left(1+(1-q)^2 r\right)^{-1/2}$ and a phase diagram over verification pressure $\Lambda$ and signal-to-noise ratio $r$. It proves a collapse condition $ (1-q)\gamma \ge f'(0)$ under which truthful effort vanishes, producing a systemic drift to proxy sovereignty unless verification bandwidth is expanded. The authors translate these results into concrete verification-first AI design guidelines (reducing claim rate $R$, increasing high-fidelity evidence to lower $C_{\mathrm{eff}}$, expanding $\mathcal{B}_{\mathrm{eff}}$, and dampening proxy drift) and outline an empirical agenda. The work advocates AI as an adversarial auditor that generates auditable evidence to extend verification bandwidth, with practical implications for venues, tool builders, and funders.

Abstract

This paper argues that AI-assisted peer review should be verification-first rather than review-mimicking. We propose truth-coupling, i.e. how tightly venue scores track latent scientific truth, as the right objective for review tools. We formalize two forces that drive a phase transition toward proxy-sovereign evaluation: verification pressure, when claims outpace verification capacity, and signal shrinkage, when real improvements become hard to separate from noise. In a minimal model that mixes occasional high-fidelity checks with frequent proxy judgment, we derive an explicit coupling law and an incentive-collapse condition under which rational effort shifts from truth-seeking to proxy optimization, even when current decisions still appear reliable. These results motivate actions for tool builders and program chairs: deploy AI as an adversarial auditor that generates auditable verification artifacts and expands effective verification bandwidth, rather than as a score predictor that amplifies claim inflation.

Figures (3)

• Figure 1: Truth-coupling phase diagram. Contours of $\rho=\mathrm{Corr}(S,T)$ over verification pressure $\Lambda$ and noise-to-signal ratio $r$ (Theorem \ref{['thm:tc']}). For $\Lambda\le 1$ we have full verification ($q=1$) and thus $\rho=1$. For $\Lambda>1$, truth-coupling degrades as verification becomes scarce (larger $\Lambda$) or proxy noise increases (larger $r$).
• Figure 2: Incentive collapse for truth effort. The researcher's optimal truth effort $e^\star$ as a function of the effective proxy return $(1-q)\gamma$ (Theorem \ref{['thm:effort']}). As $(1-q)\gamma$ increases, $e^\star$ decreases and eventually hits the corner solution $e^\star=0$ when $(1-q)\gamma \ge f'(0)$.
• Figure 3: Imitation AI automates the appearance of peer review, optimizing for style match. Verification-first AI automates the labor of scientific scrutiny, optimizing for truth discovery.

Theorems & Definitions (20)

• Definition 3.2: Truth-coupling
• Theorem 4.1: Truth-coupling under mixture evaluation
• Corollary 4.2: Coupling budget
• Theorem 4.4: Truth effort decreases with verification scarcity
• proof
• proof
• proof
• Proposition 4.2: Typical scale of best-of-$K$ grows polynomially
• Corollary 4.3: A route to exponential growth in proxy-selection noise
• Corollary 6.2: Citation truth-coupling