AI Submission Growth Needs a Capacity Policy

A manuscript queue can fail quietly. Nothing dramatic has to happen. The submission count rises, deputy editors spend more evenings on initial screens, reviewers decline a little faster, reports get thinner, and the journal still appears to be working because decisions keep going out. By the time leadership notices, the system has already normalized a worse version of itself.

That is why the recent Organization Science analysis deserves attention beyond business schools. In "More Versus Better: Artificial Intelligence, Incentives, and the Emerging Crisis in Peer Review," published online April 27, 2026, Claudine Gartenberg, Sharique Hasan, Alex Murray, and Lamar Pierce examined 6,957 initial submissions and 10,389 text-entered reviews handled by the journal since 2021: https://pubsonline.informs.org/doi/10.1287/orsc.2026.ed.v37.n3. They report that submission volume rose 42% after the late-2022 release of ChatGPT, while writing quality declined, with AI-generated writing accounting for nearly all of those trends.

A Washington University summary of the work notes that most of the additional submissions were rejected, many at the deputy-editor screen, and that more than 30% of reviews showed some degree of AI use: https://source.washu.edu/2026/05/analysis-reveals-ais-impact-on-research-journals/. The Financial Times brought the same concern into broader public view this week, framing it as a question of research volume, quality, and peer review capacity: https://www.ft.com/content/52e688a0-c6c1-4161-9878-1fab12c5e806.

The useful lesson is not that every journal should copy Organization Science or that every AI-assisted paper is weak. The authors themselves warn against treating detection as an individual verdict. The lesson is operational: AI has reduced the cost of producing plausible submissions faster than journals have redesigned the scarce human judgment that evaluates them.

Do Not Turn A Capacity Problem Into A Detector Verdict

The tempting response is to buy or build a stronger AI detector and send manuscripts with high scores to a rejection lane. That may feel decisive, but it confuses two tasks. Aggregate detection can help a journal understand how the queue is changing. Individual detection is much harder to use fairly because language background, editing support, translation, discipline style, and model evolution all affect the signal.

Organization Science handled this distinction carefully. Its analysis used scores to study large-scale trends, not to define an acceptable percentage of AI use for a specific author. Journal leaders should keep that boundary. A dashboard that shows rising AI-written abstract patterns can justify more triage staffing, clearer desk-reject rules, or revised author instructions. It should not become a black box that ends a manuscript without editorial judgment.

This matters especially for multilingual author communities. AI can be a legitimate language-support tool. A journal that treats all AI signal as misconduct may punish authors who are using tools to make real work readable. A journal that ignores the signal entirely may flood editors with thin, synthetic, or underdeveloped papers. The governance question is not "AI or no AI." It is whether the journal can distinguish assistance, overproduction, and manipulation without outsourcing judgment to a score.

The First Decision Needs More Editorial Design

Desk rejection used to be treated as an editorial art: experienced editors knew what belonged in the journal and what did not. That judgment still matters, but volume pressure makes undocumented art brittle. If submissions rise by even 20%, the first decision becomes a capacity control. At 42%, it becomes a governance control.

A good initial screen now needs explicit lanes. One lane is scope: the manuscript plainly does or does not belong in the journal. Another is readiness: the question may fit, but the paper lacks enough methodological, evidentiary, reporting, or citation substance to send to reviewers. A third is integrity risk: image concerns, implausible references, paper-mill signals, suspicious reviewer suggestions, or undisclosed reuse require escalation before normal review. A fourth is language and presentation: the work may be sound, but the paper needs author-side repair before peer review would be a responsible use of reviewer time.

Those lanes should produce different letters. "Out of scope" is not the same as "not ready for review." "Language support needed" is not the same as "integrity concern." If the journal sends the same generic desk-reject note for all four, authors learn nothing, staff cannot audit patterns, and editors lose the ability to defend the queue design when rejection rates climb.

Low-Cost Rejection Is Not Low-Quality Review

The phrase "low-cost rejection" can sound hostile to authors. It should not. A fast, specific, well-governed initial decision is often kinder than a slow review process that consumes two reviewers and returns the same conclusion eight weeks later. The ethical problem is not speed. It is opacity, inconsistency, and lack of accountability.

Journal leaders can set standards for fast decisions without pretending every screened paper receives full review. The editor should record the primary reason, select from controlled categories, add a short free-text note when needed, and preserve any integrity escalation separately from ordinary quality assessment. The appeal route should be narrow but real: authors should know what kind of new information could change the decision and what kind will not.

This is where capacity policy protects fairness. Without it, two deputy editors may treat similar AI-polished but weak manuscripts differently. One may desk reject for poor contribution, another for writing quality, another for suspected AI use, and another may send it to review because the topic is fashionable. The journal then has a hidden policy made out of individual fatigue.

The Reviewer Pool Cannot Be The Shock Absorber

Reviewers are already being asked to absorb too many institutional failures: unclear scope, weak screening, special-issue pressure, author overproduction, and now AI-assisted manuscript volume. Sending more marginal papers to reviewers is not neutral. It taxes the very people whose careful judgment distinguishes a journal from a posting service.

Reviewer-side AI use adds another layer. Wiley's current publishing ethics guidance says reviewers may use AI technology to improve the clarity or quality of written feedback only with disclosure to the handling editor, and says editors and peer reviewers must not upload manuscripts, figures, or tables into AI tools: https://authors.wiley.com/ethics-guidelines/index.html. Whether a journal follows Wiley's exact model or another publisher policy, the operational need is the same: reviewers need a clear confidentiality rule, and editors need a way to record whether AI was used in report preparation.

The queue design should therefore protect reviewers before invitation. Do not ask them to rescue manuscripts that should have been screened out. Do not make them interpret AI policy from scattered guidance. Do not treat a reviewer report as usable if it appears to have delegated confidential analysis to an external tool in violation of policy. A journal that cannot enforce those boundaries will spend its reviewer goodwill on avoidable work.

What To Measure Before The Queue Breaks

• Monthly initial submissions by article type, region, special issue, transfer source, and author history, compared with editor capacity.
• Desk-reject rate by reason code, not only overall rejection rate.
• Median time from submission to first screen, and the share of papers waiting longer than the journal standard.
• Reviewer invitations per externally reviewed paper, acceptance rate, decline reasons, and repeat load on the most-used reviewers.
• Integrity escalations at intake, including reference problems, image concerns, reviewer-suggestion anomalies, and undisclosed reuse.
• Appeal volume and appeal outcomes by first-decision category.
• Evidence that language-support cases are separated from suspected manipulation or low scholarly contribution.

These are not vanity metrics. They tell leadership whether AI-era volume is being absorbed by better workflow or by hidden labor. If deputy editors are spending twice as much time on first screens, if reviewers are declining faster, or if appeals cluster around one vague rejection category, the journal has a design problem.

A Capacity Policy Authors Can Read

The author-facing policy should be plain. It should say that AI-assisted language and editing may be allowed where disclosed and verified by the authors, that authors remain responsible for arguments, evidence, citations, and analysis, and that manuscripts may be returned or rejected before review when they are not sufficiently developed for expert evaluation. It should also say that suspected manipulation, fabricated references, or attempts to game review will be handled as integrity concerns, not as writing-quality issues.

Internally, the policy should assign authority. Who can desk reject for readiness? Who can require resubmission before review? Who escalates possible manipulation? Who reviews appeals? Who can override an AI-related concern? Who audits whether the rules are affecting some author groups disproportionately? If those questions are unanswered, the policy will collapse into editor preference under load.

The point is not to make authors afraid of tools. It is to make clear that the journal evaluates research, not fluent packaging. AI can help authors edit, translate, search, code, or structure ideas. It cannot be allowed to turn peer review into a volume game where the cheapest manuscript to generate becomes the most expensive one for the community to evaluate.

The Takeaway For Journal Leaders

Treat AI-driven submission growth as a capacity and governance issue before treating it as a misconduct issue. Build a first-screen policy with distinct decision lanes, measure the pressure on editors and reviewers, keep detector data at the aggregate level unless supported by human assessment, and give authors readable rules for what happens before peer review.

The journals that handle this well will not be the ones with the sternest AI paragraph. They will be the ones that can show, month by month, that scarce expert review is being reserved for manuscripts ready to benefit from it.