Fighting Fire with Fire: Infusing Artificial Intelligence into Peer Review to Sustain Quality Scholarship

1. Introduction and Motivation

Peer review is essential to academic research and the profession. It requires multiple expert reviewers with specialized knowledge to devote substantial time in assessing the quality of submitted articles. As with any operations system, peer review faces substantial stress when the supply of submissions overcomes the capacity to process them. Such a scenario is occurring now as artificial intelligence (AI) tools, especially generative AI and large language models (LLMs), have accelerated authors’ ability to generate research articles. For instance, Google’s AI Co-Scientist has successfully generated novel, literature-grounded hypotheses for biology laboratories, some of which aligned with hypotheses developed independently by the same laboratories (Gottweis et al. 2025). Similarly, Gans (2025) reports that an AI coauthor helped draft a full paper within a single day. Mathematician Terence Tao has likewise explored AI collaboration on proofs and analyses (Tao 2025). Ludwig and Mullainathan (2024) and Manning et al. (2024) indicate that AI holds potential for generating and testing hypotheses within the social sciences.

As a result, AI appears to make authors more productive. Kusumegi et al. (2025) provide mounting evidence that LLMs boost the output of scientific papers by 23.7%–89.3% depending on the field and the author’s background. Xu and Yang (2026) show that an agentic AI workflow can reanalyze and replicate studies with a high success rate conditional on accessible data and code, which should raise author productivity by letting researchers produce, verify, and revise more work in less time. More recently, Project Autonomous Policy Evaluation (APE)¹ illustrated the impact of this productivity gain in practice; its public dashboard reports more than 400 AI-generated economics papers produced within two months. Unlike the studies above, which focus on AI reshaping individual researchers’ workflow, APE demonstrates what happens when AI-assisted authorship is deployed as a systematic pipeline; the paper output volume of a small team begins to drastically scale.

A recent editorial commentary in Information Systems Research underscores that generative AI is already reshaping research practices and disciplinary expectations while warning that the community must develop new norms to manage these shifts responsibly (Gopal et al. 2025). From a recent report by a top publisher: “Peer review … is under unprecedented pressure. Too many submissions are funnelled through a limited pool of reviewers, creating unsustainable workloads and threatening the effectiveness, speed and fairness of peer review” (Cambridge University Press 2025, p. 11). This article articulates the operations crisis confronting academic peer review, develops a formal model to capture the essence of the process, proposes an AI-infused peer review workflow, and applies an extended model to examine under what conditions and to what extent this new approach can help address the crisis (Cambridge University Press 2025). Our process and model exemplify how AI can be integrated, but we recognize that designing the optimal way to employ AI in the review process requires careful design, calibration, and experimentation.

1.1. Impending Crisis: Operational Strain and the “Shadow AI” Problem

As AI increases submissions, academic journals are finding it harder to distinguish high-quality work while maintaining timely decisions (Gartenberg et al. 2026). Management Science submissions increased from 4,710 in 2024 to 5,735 in 2025, a rise of 22%. At two major journals, submissions increased by roughly 50% between 2022 and 2025 (panel (a) of Figure 1). Recent concerns on manuscript and grant time to decisions have been noted by government agencies and in the academic press (National Institutes of Health 2025, Richardson et al. 2025). Operations theory suggests a nonlinear congestion effect; as utilization approaches one, small inflow shocks translate into large delay increases—a capacity cliff. Panel (b) of Figure 1 illustrates this pattern based on a stylized operations model of journals (presented in Section 2), with nonlinear deterioration in the critical metrics that define a journal’s performance, namely turnaround time and decision accuracy.

A related problem is shadow AI, where reviewers (and editors) already use AI tools but in an opaque and ungoverned manner. Although several academic journals and publishers restrict these AI tools in peer review because of concerns over confidentiality, accountability, and accuracy (Bhargava et al. 2025, Peters and Chin-Yee 2025), informal reports and detection-based studies have documented that a nontrivial fraction of reviews (in computer science conferences) contain AI-generated text (Liang et al. 2024a, Latona et al. 2025). For journal leadership, shadow AI is difficult to observe or measure. When multiple reviewers privately rely on similar general-purpose LLMs, their review errors become structurally correlated, eroding the independence of judgment that journals rely upon. Moreover, reviewer-initiated use of public AI interfaces can create uncontrolled data retention, use a variety of unknown prompts, or have third-party logging. A governed, journal-specific AI system does not eliminate shadow AI, but it may reduce reviewers’ incentive to seek external AI support while making the journal’s own AI use transparent and its induced correlations estimable.

1.2. Proposed Solution: Managed, AI-Augmented Peer Review

Although AI use for peer review is problematic when it is opaque, uncontrolled, heterogeneous, and undisclosed, there is credible evidence of an upside to using AI. Comparative studies report substantial overlap between LLM critiques and human critiques (Liang et al. 2024b). They can be helpful in identifying clarity problems, missing reporting elements, internal inconsistencies, and other structured checks (Berente and Recker 2025).

The managerial problem is, therefore, not “AI or no AI” but how to move from opaque adoption to managed adoption of AI: a journal-provided, journal-scoped AI diagnostic tool applied before human review paired with a short author right of response and instrumented through structured logging. The AI output is explicitly nonfinal; it is an input to editorial judgment, not a recommendation and not a substitute for reviewers. We propose that the AI reviewer and process be journal specific. General-purpose AI tools lack alignment with a particular journal’s standards, methodologies, and values. Therefore, journals will scope AI to tasks where criteria can be specified and outputs can be validated, and they may prevent scope creep into value judgments.

The AI tool is a design choice that can be tailored, trained, and vetted according to the specific criteria of the target journal, potentially focusing its capabilities on aspects appropriate for that journal (e.g., mechanical checks, writing quality, or methodological soundness) while freeing up resources for human oversight and judgment. Journals must customize their AI tool; some journals may constrain the AI to auditable rubrics, and others may allow for less constrained AI reviews. Confidentiality and intellectual property concerns further strengthen the case for a journal-led workflow.

Managed AI (versus both shadow AI and permissive policies) enables the measurement of the effects of using AI. It enables treatment of AI as an informational input with two estimable properties: reliability (noise/precision of the AI signal) and error correlation (the degree to which review errors are correlated with each other once the AI review is observed). These quantities could in principle be recovered from records or pilot rollouts and used to design workflows that meet a journal’s standards. Section 3 presents a proposed workflow, which we formalize by extending our journal review model by infusing AI into the process. Given a manuscript to review, humans (or humans and AI in the governed-AI case) produce noisy unbiased estimates of manuscript quality, and the journal seeks precise estimates to improve its decision accuracy. The AI review (and initial author response) affects two channels: the productivity of reviewer time and the weight that the AI review carries in the reviewer’s final report. The risk is error correlation; when reviewers see the AI report, their errors may become correlated with the AI’s errors (anchoring).

2. A Benchmark Model of Journal Operations

We frame an academic journal’s review process as a resource-constrained estimation problem. The journal receives a flow of submissions and through editorial review, must estimate manuscript quality accurately enough to make sound accept/reject decisions (see Figure 2).

The key performance dimension is estimation-error variance: how precisely the deciding editor (DE) can assess manuscript quality given limited reviewer time. Improved precision allows the journal to more often correctly accept good papers and reject weak ones. As mentioned earlier in Section 1.1, as submission rates rise—particularly with AI-accelerated authorship—the reviewer time available per manuscript shrinks, putting upward pressure on estimation-error variance and threatening decision quality.

Formally, we model journal review as a stylized single-round process. We assume that manuscript quality is a scalar latent variable Q. Each manuscript is assigned to $m^H$ human reviewers, where H denotes human parameters and outcomes in the human-only benchmark case. Each reviewer i provides a quality signal ${\widehat{Q}}_i^H$ to a deciding editor. Reviewer signals are noisy, with precision improving in reviewer skill and time allocated.² The DE aggregates the reviewer signals to form an estimate of manuscript quality, and the journal sets a target estimation-error variance ${\sigma }_{\text{tar}}^2$ that determines the required review thoroughness. The DE’s only choice variable is the time allocated per reviewer, t; all other parameters are taken as given.

This parsimonious setup blends multiround review into one stage so that we can focus on estimation precision rather than iterative paper improvement. A valuable extension may be to consider improvements to paper quality endogenous to the review process. The notation is summarized in Table 1.

Table 1. Notation Summary

2.1. Human Reviewer Signals and Aggregation

Reviewer i produces the signal

where $t^H>0$ is the time allocated to each reviewer and ${\tau }^H(t^H)$ is the time-effect function, which is increasing and concave: ${({\tau }^H)}^{\prime }(t^H)>0$, ${({\tau }^H)}^{″}(t^H)\le 0$, ${\tau }^H(0)=0$, and ${\tau }^H(\infty )\to 1$. The time-effect function translates time invested by the reviewer into increasing review precision (by reducing ${\epsilon }_i^H$). Although this function may be sigmoidal, we assume that the journal operates in the concave portion; hence, ${\tau }^H(t^H)$ is concave in $t^H$. Note that although ${\widehat{Q}}_i^H=S_i^H$ in this benchmark, this will not be the case in the AI-assisted regime.

Assume that reviewer errors ${\{{\epsilon }_i^H\}}_{i=1}^{m^H}$ are conditionally independent given Q. The human-panel mean and estimation-error variance in the human-only benchmark is

The variance is decreasing in both $t^H$ and $m^H$ (with diminishing incremental gains from additional reviewers (see the Online Appendix). A smaller estimation-error variance ${\sigma }_{\text{post}}^{2,H}$ enables more accurate acceptance decisions. Given a target estimation-error variance ${\sigma }_{\text{tar}}^2>0$, the DE’s optimal (minimal) choice of per-reviewer time is $t^{*,H}={\left({\tau }^H\right)}^{-1}(1/(m^H{\theta }_H{\sigma }_{\text{tar}}^2\left)\right)$.

2.2. Throughput and the Capacity Cliff

Given a total reviewer-hour budget $H_{\text{tot}}$ per period, the journal’s service rate (throughput) is ${\mu }^H≔H_{\text{tot}}/(m^H\hspace{0.17em}{t}^{*,H})$. Manuscripts arrive as a Poisson process with rate $\lambda >0$. By standard results from queueing theory, the expected journal response time $T^H(\lambda ;{\mu }^H)=1/({\mu }^H-\lambda )$ is strictly increasing and convex in $\lambda$, and it diverges as $\lambda \uparrow {\mu }^H$. This capacity cliff property means that as AI-accelerated authorship drives up submissions, journals operating near their human-only capacity ${\mu }^H$ will exhibit longer turnaround times. This puts downward pressure on decision accuracy (i.e., upward pressure on estimation-error variance, ${\sigma }_{\text{post}}^{2,H}$) as journal reviewers spend less time per submission to keep turnaround times under control even as submission volume increases.

Figure 1 from Section 1 illustrates this emerging crisis. At relatively low submission volumes, journals comfortably maintain high decision accuracy with short turnaround times. As the submission rate grows, the system reaches the capacity cliff; reviewer workloads consume the journal’s capacity, leading to rapidly escalating turnaround times and processing backlogs. Journals face the difficult choice of increasing reviewer resources, reducing the number of reviewers per manuscript, or accepting lower decision accuracy.

3. Fighting Fire with Fire: AI-Infused Peer Review

In its recent review of operational pressures facing academic journals, the Cambridge University Press underlines the role of technology in improving productivity of peer review (Cambridge University Press 2025). Calling for “new solutions that address peer review at scale,” the review argues that “responsible use of AI [can] streamline administrative and technical aspects of review, allowing human expertise to focus on what makes research interesting and important” (Cambridge University Press 2025, p. 12). We propose an AI-infused hybrid editorial workflow, which strategically incorporates AI early in the review process, potentially reconfiguring the efforts of human reviewers more toward high-level critical tasks. AI influence is explicit and auditable. There is no intent to automate editorial judgment; instead, our proposed approach suggests leveraging AI to support human judgment with the aim to improve the sustainability of the peer review process. We recognize that identifying an “optimal” AI-infused peer review workflow or even definitively projecting the consequences of one will require substantial experimentation to calibrate and configure the workflow across multiple alternative designs.

3.1. Proposed AI-Augmented Journal Operations Workflow

The core of our proposal is a two-stage process in which the evaluative role of AI is tailored to the domain, methods, and value systems of the journal.

1. Initial AI review and author response. Upon submission, every paper undergoes an immediate, journal-specific AI review. The AI tool generates structured critique and feedback for the authors, who then prepare a response document (or withdraw the paper) to address or rebut the points raised by the AI. This initial exchange serves to identify and correct basic issues, increases clarity, and ensures a baseline level of quality before the paper enters the human review process.

2. Human review with AI context. When the paper proceeds, human reviewers receive not only the original manuscript but also, the AI-generated review and the authors’ response. This augmented context allows them to save time on mechanical checks, which now simply need human oversight, and focus their intellectual energy on aspects of the review that require deeper human involvement, such as contribution to the literature and impact.³

Our proposal specifically implies that the AI review focuses on activities that have been vetted with respect to that journal. For instance, a particular journal might use AI to examine writing quality or a rigorous methodology that provides evidence of these claims.⁴ Crucially, this proposal shifts the use of AI from an informal, unobserved tool used idiosyncratically by individuals to a journal-provided and journal-scoped assessment that every submission sees in the same way. The AI and human reviewers work together as explained below and depicted in Figure 3. The manuscript itself remains fixed at this stage; the response is an attachment, not a revision. Reviewers receive three objects: the manuscript, the AI assessment, and the author response.

3.2. A Model of AI-Infused Peer Review

We extend the model of peer review to capture the AI-infused peer review process described in Figure 3. The model captures two key tensions. (1) AI with quality ${\theta }_Z$ can make reviewer time more productive (captured through ${\tau }^A$) and improve the precision of human reviews through an informative signal $Z({\theta }_Z)$ comprising the AI review and author response, but (2) it can reduce the diversification benefit of having multiple independent reviewers because all reviewers rely on the same AI review (through the weighting term $\omega ({\theta }_Z)$), creating potential correlation ${\rho }_{HH}$ in their errors.

The AI review affects the editorial decision only indirectly through its effect on the human reviewers’ reports. Our model’s focus is on reviewer time. We do not model DE time per review nor model authors’ withdrawals or desk rejections. We also do not model the author response to the AI review, which adds elapsed time to the process but not to reviewer time. Lastly, we do not model the potential effect of AI-assisted review on the mean number of review rounds of a multiround process.

3.2.1. AI Signal.

Let the AI review signal be $Z({\theta }_Z)=Q+{\epsilon }^Z$ with unbiased mean $\mathbb{E}[{\epsilon }^Z]=0$ and variance $\text{Var}({\epsilon }^Z)={\sigma }^{2,Z}({\theta }_Z)$, where ${({\sigma }^{2,Z})}^{\prime }({\theta }_Z)<0$. Without loss of generality (because of the freedom in defining the scale of ${\theta }_Z$), set ${\sigma }^{2,Z}({\theta }_Z)=1/{\theta }_Z$.

3.2.2. Human Reviewers in the AI-Assisted Regime.

Let $t^A>0$ denote the time spent by each of the $m^A$ reviewers in the AI-assisted regime. Reviewer i’s assessment ${\widehat{Q}}_i^A$ can be viewed as a convex combination of the reviewer’s own assessment $S_i^A$ plus the AI review given to them:⁵

where $\omega ({\theta }_Z)\in [0,1)$ measures the weight placed on the AI review (treated as exogenous). Note that ${\theta }_H$ appears in both regimes because the human reviewer’s intrinsic ability is regime invariant. In the AI-assisted regime, the reviewer’s signal $S_i^A$ uses the ${\tau }^A$ time-effect function, which we write as ${\tau }^A(t^A,{\theta }_Z)>0$ (i.e., we capture dependence of time t and AI quality ${\theta }_Z$).⁶

We assume mutual independence of intrinsic reviewer errors ($\text{Cov}({\epsilon }_i^A,{\epsilon }_j^A)=0$ for $i\ne j$) and between the reviewer and AI errors $\text{Cov}({\epsilon }_i^A,{\epsilon }^Z)=0$. We also assume that there is no additional distortion term induced by AI exposure. AI and reviewer signals are treated as unbiased estimates of manuscript quality ($\mathbb{E}[{\epsilon }^Z]=\mathbb{E}[{\epsilon }_i^A]=0$). The model captures only the effect of AI-assisted review on the variance channel and does not capture the directional bias (mean-effect) channel.

Under these assumptions, AI can worsen the editorial decision through two channels. (i) The shared AI error ${\epsilon }^Z$ introduces between-reviewer error correlation that reduces the diversification benefit of multiple reviewers, and (ii) the time-effect function ${\tau }^A(t^A,{\theta }_Z)$ may be lower than ${\tau }^H(t^H)$ if the AI review and author response prove distracting or misleading.

3.2.3. Induced Covariance Across Human Reviewers.

Because all reviewers observe the same AI review, the errors of their final report become correlated through the shared AI error term ${\epsilon }^Z$; the signals remain unbiased. For a given reviewer i,

For two distinct reviewers $i\ne j$, $\text{Cov}({\widehat{Q}}_i^A-Q,\hspace{0.17em}{\widehat{Q}}_j^A-Q)=\omega {({\theta }_Z)}^2\hspace{0.17em}{\sigma }^{2,Z}({\theta }_Z)$.

The pair-wise between-reviewer error correlation is

This error correlation is generated within the workflow itself because all reviewers see the same AI review. (The formal derivation is in the Online Appendix.) The error correlation ${\rho }_{HH}$ is governed by the relative magnitude of shared versus idiosyncratic error variance. When $\omega ({\theta }_Z)$ is small, reviewers rely primarily on their own individual assessments and ${\rho }_{HH}\approx 0$. As $\omega ({\theta }_Z)\to 1$, ${\rho }_{HH}\to 1$.

3.2.3.1. Final Signal to DE.

This is the observed panel mean

where ${\overline{\epsilon }}^A≔(1/m^A){\sum }_{i=1}^{m^A}{\epsilon }_i^A$. The DE’s estimation-error variance is

Online Appendix B.4 shows that the first term is a floor for the nondiversifiable AI error variance and invariant to $m^A$ and $t^A$; the second term, the diversifiable human error variance, shrinks with $m^A$ and $t^A$.

3.2.4. Comparative Statics with Respect to AI Quality.

The effect of AI quality on the estimation-error variance is not unambiguously beneficial. Differentiating ${\sigma }_{\text{post}}^{2,A}(t^A,{\theta }_Z)$ with respect to ${\theta }_Z$, the sign depends on three forces:

1. the effect of AI quality on the time-effect function through $\partial {\tau }^A/\partial {\theta }_Z$,

2. the direct improvement in AI precision through ${({\sigma }^{2,Z})}^{\prime }({\theta }_Z)<0$, and

3. the weighting on the AI signal through ${\omega }^{\prime }({\theta }_Z)$.

Force (1) is ambiguous in sign; better AI and author response may help reviewers more effectively review the paper ($\partial {\tau }^A/\partial {\theta }_Z>0$), or the additional complexity may hinder their review effectiveness ($\partial {\tau }^A/\partial {\theta }_Z<0$). Force (2) is unambiguously beneficial, with higher AI quality providing a more accurate AI signal $Z({\theta }_Z)$. Force (3) creates a tension; higher weight on the AI signal amplifies the nondiversifiable common error but reduces diversifiable idiosyncratic noise. Better AI improves estimation precision when the direct precision gain (force (2)) and any time-productivity gains (force (1)) dominate the increased error correlation from higher weighting (force (3)). (A formal necessary and sufficient condition is stated in the Online Appendix.)

3.3. Main Analytical Results

We first record a governance result—that optimal reliance on the AI review is interior (Proposition 1)—and then, establish two threshold results that characterize when AI quality is sufficient to deliver concrete resource savings. The first is an intensive-margin threshold: the AI quality above which better AI reduces the per-reviewer time needed to achieve a fixed estimation-error target. The second is an extensive-margin threshold: the AI quality above which the journal can achieve the same quality target with fewer human reviewers. (Formal proofs are in Online Appendix B.)

Proposition 1

(Optimal Reliance on AI). Fix the number of reviewers $m^A$, per-reviewer time $t^A$, and AI quality ${\theta }_Z$, and write the AI signal noise as $A≔{\sigma }^{2,Z}({\theta }_Z)$ and the human noise as $B^A≔1/\left[m^A{\theta }_H{\tau }^A\right(t^A,{\theta }_Z\left)\right]$. The DE’s estimation-error variance as a function of reviewer reliance $\omega \in [0,1)$, ${\sigma }_{\text{post}}^{2,A}(\omega )={\omega }^{2}A+{(1-\omega )}^2{B}^A$, is strictly convex and minimized at the interior weight ${\omega }^{*}=B^A/(A+B^A)\in (0,1)$. Overreliance ($\omega >{\omega }^{*}$) inflates the common AI-induced error and the crossreviewer correlation ${\rho }_{HH}$; underreliance forgoes the informative AI signal.

Proposition 1 characterizes the variance-minimizing weight. This is increasing in quality ${\theta }_Z$, holding the time-effect channel ${\tau }^A$ fixed. It is a benchmark, not a weight set directly. The condition under which AI-assisted review beats the human-only panel and its proof are given as Proposition 5 in Online Appendix D.3.

3.3.1. Target Variance and Required Reviewer Time.

Let ${\sigma }_{\text{tar}}^2>0$ be the DE’s target estimation-error variance. Define the DE’s optimal time allocation in the AI-assisted regime $t^{*,A}({\theta }_Z)$ implicitly by ${\sigma }_{\text{post}}^{2,A}(t^{*,A}({\theta }_Z),{\theta }_Z)={\sigma }_{\text{tar}}^2$.

The following observation records the intensive-margin implication; when AI quality reduces estimation-error variance at the operating point, the reviewer time required to hit a fixed target also falls.

Observation 1

(Effort Saving). Suppose that (1) $\partial {\sigma }_{\text{post}}^{2,A}/\partial t^A<0$ and (2) $\partial {\sigma }_{\text{post}}^{2,A}/\partial {\theta }_Z<0$ evaluated at $(t^A,{\theta }_Z)=(t^{*,A}({\theta }_Z),{\theta }_Z)$. Then, $\partial t^{*,A}\left({\theta }_Z\right)/\partial {\theta }_Z<0.$ Thus, higher AI quality reduces the reviewer time required to achieve a fixed estimation-error target.

Condition (1) in Observation 1 holds whenever $\partial {\tau }^A/\partial t^A>0$, which we assume throughout. Condition (2) in Observation 1 is the substantive requirement; it holds when the three forces identified above—direct AI precision improvement, the weight of the AI signal effect, and the time-effect function response—net out so that better AI quality reduces the estimation-error variance at the operating point where the target is just met.

3.3.2. Intuition.

The human-only benchmark requires a fixed per-reviewer time $t^{*,H}$ that does not depend on ${\theta }_Z$. When the conditions of the observation hold, $t^{*,A}({\theta }_Z)$ is strictly decreasing in ${\theta }_Z$. Because $t^{*,A}$ is continuous and decreasing, whereas $t^{*,H}$ is constant, there exists a threshold AI quality ${\widehat{\theta }}_Z$ satisfying $t^{*,A}({\widehat{\theta }}_Z)=t^{*,H}$ (provided that AI quality is high enough that $t^{*,A}$ eventually falls below $t^{*,H}$). For all ${\theta }_Z>{\widehat{\theta }}_Z$, the AI-assisted regime achieves the same estimation-error target with strictly less reviewer time per manuscript. This threshold represents the minimum AI quality at which the journal begins to realize reviewer time savings from AI integration on the intensive margin. If embedded in a queueing layer, these time savings translate into higher reviewer throughput, holding journal quality fixed. We have modeled reviewer time, but we have not modeled DE time, author response time, or the number of rounds of review. The AI review and author response step may add elapsed calendar time before human review begins; our model captures reviewer time requirements rather than total calendar latency, so this front-end latency must be evaluated empirically.

3.3.3. Reviewer Substitution Threshold.

Let the human-only benchmark consist of $m^H=3$ human reviewers with per-reviewer time $t^{*,H}$. Now, consider the AI-assisted regime with $m^A=2$ human reviewers evaluated at the benchmark reviewer time $t^{*,H}$, and define $G({\theta }_Z)≔{\sigma }_{\text{post}}^{2,A}{|}_{m^A=2}(t^{*,H},{\theta }_Z)$.

Proposition 2

(Substitution Threshold). Assume that $G({\theta }_Z)$ is continuous and strictly decreasing on an interval $[{\underset{¯}{\theta }}_Z,{\overline{\theta }}_Z]$ and that $G({\underset{¯}{\theta }}_Z)>{\sigma }_{\text{tar}}^2$ and $G({\overline{\theta }}_Z)<{\sigma }_{\text{tar}}^2$. Then, there exists a unique critical AI quality ${\theta }_Z^{\text{crit}}\in ({\underset{¯}{\theta }}_Z,{\overline{\theta }}_Z)$ such that $G({\theta }_Z^{\text{crit}})={\sigma }_{\text{tar}}^2$. For every ${\theta }_Z>{\theta }_Z^{\text{crit}}$, two AI-assisted reviewers strictly exceed the target precision previously delivered by three human reviewers without AI. Under these conditions, AI quality induces a unique substitution threshold.

The existence of the threshold is conditional on $G({\theta }_Z)$ being monotone decreasing over the relevant operating region; the previous comparative-statics discussion shows that this need not hold globally because AI quality affects the system through multiple competing channels. This threshold corresponds to the point at which one human reviewer can be removed without sacrificing accuracy. It provides an operational benchmark for when AI-assisted reviewer substitution is feasible. In the Online Appendix, we present formal derivations in Online Appendix C, an analysis of review time in a queueing layer in Online Appendix B, and an analysis of our model using closed-form functional forms in Online Appendix D.

We close this analysis with the crucial result that the AI-infused workflow can improve decision accuracy even when the stand-alone AI review signal is noisier than that of a time-constrained human reviewer (i.e., $(1/{\theta }_Z)>(1/[{\theta }_H{\tau }^H\left(t^H\right)\left]\right)$). This condition is stronger than a condition that AI reviewer quality is below human reviewer quality (${\theta }_Z<{\theta }_H$). The mechanism and intuition behind this result are that the AI review allows the journal to reduce the number of reviewers per paper, allowing each reviewer more time and hence, improving their signal to the DE. Analyzing the effect of this process change in conjunction with other aspects, such as the scoping of AI review and the degree to which human reviewers are influenced by the AI review when doing their own evaluation, yields the condition under which AI-infused review can be beneficial. The feasible region covers cases where ${\theta }_Z<{\theta }_H$. We state the result formally in Proposition 3 and provide a numerical example to illustrate the finding.

Proposition 3

(AI-Infused Review Can Be Effective Even When AI Is Weaker Than Human). Suppose the AI-assisted review regime cuts the number of human reviewers from $m^H=3$ (each with time $t^H$) to $m^A=2$ (each with time $t^A$) while holding total human reviewer time per manuscript fixed, $m^H{t}^H=m^A{t}^A$. Assume that ${\epsilon }_1^A$, ${\epsilon }_2^A$, and ${\epsilon }^Z$ are mutually independent. Then, the AI-assisted regime can yield a lower-variance signal to the DE than the human-only benchmark, even when $(1/{\theta }_Z)>(1/[{\theta }_H{\tau }^H\left(t^H\right)\left]\right)$ (i.e., the stand-alone AI review signal is noisier than that of a time-constrained human reviewer and hence, also larger than the unconstrained human; that is, $(1/{\theta }_Z)>(1/{\theta }_H)$).

Proposition 3 has been framed above as an accuracy result—lower estimation-error variance at fixed total reviewer time; however, the journal can instead convert some or all of the gain into shorter turnaround, fewer reviewer hours per manuscript, or a mix of accuracy and speed. We illustrate this with four numerical scenarios. The accuracy-gain framing of the proposition holds $m^H{t}^H=m^A{t}^A$ fixed, and this reallocation—fewer reviewers, each with more time—delivers accuracy gain only in scenarios 1–3. Scenario 4 below illustrates the alternative in which part of the gain is taken as a reduction in total reviewer time. The essential point is that the AI-assisted regime lets at least one metric (accuracy or turnaround) improve, whereas the other does not worsen.

Example 1.

Let $(m^H=3,m^A=2,\omega =\frac{1}{5},k=1)$ with human reviewer quality ${\theta }_H=1$. Let human-only reviewer time $t^H=-\log (0.3)\approx 1.204$. Then, ${\tau }^H(t^H)=1-e^{-t^H}=0.7$. The individual human review signal under this benchmark time allocation has noise $(1/0.7)=(1/\left[{\theta }_H{\tau }^H\right(t^H\left)\right])$. The estimation-error variance under the three human reviewer benchmark is ${\sigma }_{\text{post}}^{2,H}\approx 0.4762$.

Keeping total human reviewer time (for each manuscript) unchanged when moving from three human reviewers to two AI-assisted human reviewers, $m^H{t}^H=m^A{t}^A$, yields $t^A=(3/2)t^H\approx 1.806$. The stand-alone AI signal has noise $({\sigma }^{2,Z}\left({\theta }_Z\right)=1/{\theta }_Z)$, which exceeds the individual human review signal in scenarios 1, 3, and 4 below (in scenario 2, AI noise is worse only than the unconstrained human reviewer). The effect of AI review on human productivity is defined by $\psi ({\theta }_Z)$ (a value exceeding one represents a productivity gain) and impacts the time effect via ${\tau }^A(t^A,{\theta }_Z)=1-e^{-t^A\psi ({\theta }_Z)}$.

All scenarios yield that the AI-assisted estimation-error variance with two human reviewers

$${\sigma }_{\text{post}}^{2,A}=\left(\frac{{\omega }^2}{{\theta }_Z}+\frac{{(1-\omega )}^2}{m^A{\theta }_H{\tau }^A(t^A,{\theta }_Z)}\right)$$

is less than ${\sigma }_{\text{post}}^{2,H}\approx 0.4762$ (Table 2).

Table 2. Illustrative Examples of Effects for AI-Infused Peer Review

Table 2. Illustrative Examples of Effects for AI-Infused Peer Review

Scenario	${\theta }_Z$	$\psi ({\theta }_Z)$	$t^A$	$m^A{t}^A$	$1/{\theta }_Z$	${\sigma }_{\text{post}}^{2,A}$
1. (No productivity gain)	$\frac{1}{2}$	1	1.806	3.612	2.000	0.4629
2. (Improved AI quality)	$\frac{3}{4}$	1	1.806	3.612	1.333	0.4363
3. (AI + productivity boost)	0.65	1.279	1.806	3.612	1.538	0.4168
4. (Reallocation: $m^A{t}^A=0.9\hspace{0.17em}{m}^H{t}^H$)	0.65	1.279	1.625	3.251	1.538	0.4273

The two-reviewer AI-assisted process produces a lower-variance signal to the DE through a mix of effects. (i) Modest reliance on AI keeps the nondiversifiable AI noise floor small (${\omega }^2/{\theta }_Z$, ranging from 0.053 to 0.080 across the four scenarios), (ii) reducing the human panel from three reviewers to two gives reviewers more time per manuscript, and (iii) the AI review plus author response increases the productivity of that time through $\psi ({\theta }_Z)>1$ (in scenarios 3 and 4). Scenario 4 illustrates accuracy gain plus a 10% reduction in total reviewer time (shorter turnaround).

3.4. Shadow AI

Our analysis has covered an idealized peer review status quo, in which multiple human reviews produce independent signals and do so without using an AI review tool. As discussed earlier, emerging evidence of shadow AI in the review process suggests that reality may be different than the human-only benchmark; with ad hoc use of AI, human reviewers also become increasingly correlated with each other as they mutually (yet privately) rely on AI-assisted review tools. In our model, this corresponds to (1) unknown reviewer weighting of the AI signal and (2) private AI signals that are generated with unknown prompts by the individual reviewer.

We discuss the potential implications of shadow AI for our benchmark (a panel of independent reviewers). Reviewers using shadow AI may be more productive but may also induce a positive correlation among the human reports; reviewers privately relying on similar AI tools may share a common, journal-unobserved error component. This is similar to the common error—the induced correlation ${\rho }_{HH}$—that the governed workflow introduces, with one key difference; under shadow AI, the common component is unmanaged and unmeasured, whereas under the governed workflow, it may be observable. The AI-induced correlation that our model stresses is a cost relative to a panel of independent reviewers, but it can potentially be a benefit relative to shadow AI because the journal can observe the common component. A further distinction is the structured author response, which can meaningfully interact with the AI review in a way that shadow AI does not.

Compared with shadow AI, our proposed governed framework has three advantages. First, the AI review is produced by the journal and shown to reviewers. This makes the AI signal $Z({\theta }_Z)$ observable, and it enables the journal to estimate, perhaps through experiments, reviewers’ dependence on the AI review ($\omega$). Second, the AI signal is generated by the journal, increasing transparency to authors on what signal $Z({\theta }_Z)$ the reviewers are using. Third, the governed workflow elicits an author response to the AI review, which enters the time-effect function ${\tau }^A(t^A,{\theta }_Z)$. This response can improve reviewer productivity—an important channel that is absent under shadow AI.

3.5. Key Considerations About AI-Infused Peer Review

One concern for an adverse effect from using an AI reviewer (and sharing the AI review with human reviewers) is error correlation between human reviewers and AI output—specifically, reviewer complacency. However, this risk exists to some extent even when reviewers independently use AI tools. Observing the journal-approved AI makes this transparent, with correlation estimable potentially via experiment. To mitigate anchoring, our view is that AI output should be (i) prohibited from accept/reject recommendations and (ii) presented in structured and tunable formats (e.g., flags first with expandable evidence). Humans retain responsibility for all decisions, including contribution, novelty, impact, and ultimately, whether to accept or reject. Additional concerns involving implementation and the need for experimentation are summarized in Table 3.⁷

Table 3. Key Questions for AI-Augmented Peer Review

4. Conclusion

This article emerged from the imperative to address operational stress faced by journals from increased submission volume combined with the potential harms from hidden uncontrolled use of AI in peer review. Rather than responding defensively or resisting AI, the academic community has the opportunity to lead in shaping systems that uphold scholarly values in this new era. This paper offers both a theoretical model and a practical workflow to integrate AI into peer review (laid out in Section 3), not as a substitute for human judgment but as a structured and transparent workflow. We propose employing AI as a first-line reviewer to provide early feedback and request author response prior to human evaluation. Authors receive quick feedback from an audited, transparent, and journal-approved process, with an opportunity to clarify and explain. The AI review combined with author-provided explanation can provide greater context and focus to reviewers. A governed AI process can serve accountability requirements (human reviewers make final decisions, consistent with the European Union AI Act mandates and similar regulation), maintain confidentiality through secure infrastructure with reviewer-level access controls, and enable measurement of the process. Our framework preserves reviewer agency, supports editorial judgment, and converts unobserved processes—such as reviewers privately using their own AI tools and prompts—into managed operational levers under editorial oversight.

Our specific proposal for countering the AI-induced operational challenge is a call to action that aims to prepare journals for the realities of AI-augmented scholarship rather than a finalization in each aspect. Indeed, the final workflow will be a function of various journal-specific discussions and iterations with stakeholders. It has both merits and limitations, underlining the need for careful experimentation. There are many remaining questions and concerns for future assessment, including those highlighted in Table 3 and the appendix. We hope that this proposal establishes a foundation for the management sciences scholarly community to build upon and improve.