Researchers propose AI-first peer review, with authors responding before humans weigh in
A mathematical model suggests journals could make better decisions with less reviewer time. The risk is that a shared AI assessment makes human opinions less independent.
A proposal by Zhe Zhang and colleagues would make a journal-controlled AI assessment an explicit stage of peer review, with authors able to respond before human reviewers evaluate the paper. Mathematical modeling suggests the workflow could improve decision accuracy and reduce reviewers’ workload under some conditions, but those gains are not established in working journals and may disappear if reviewers rely too heavily on the AI. A University of Arizona workshop is testing a version of the process.
01
Elsevier submissions cited by UC San Diego rose from 2.7 million in 2022 to 4.2 million in 2025, a 56% increase.
02
Human reviewers would see the manuscript, the shared AI assessment, and the authors’ response, which could address errors before the review begins.
03
A common AI assessment could anchor reviewers and produce less independent feedback; the proposal’s modeling does not establish how often this would happen in practice.
AI could help journals make better publication decisions while taking less time from human reviewers—even if it is a weaker reviewer itself. That is the conditional finding behind a proposal Zhe Zhang explained in a November 19 UC San Diego interview. The catch: showing everyone the same AI assessment could undermine their independent judgment.
Journals would control the first assessment
Zhang, an assistant professor at UC San Diego’s Rady School of Management, and colleagues outline the approach in Management Science. Rather than leaving individual reviewers to choose their own AI tools, journals would use a customized AI reviewer, tailored to their standards, to assess manuscripts upon submission.
The workload problem is substantial. UC San Diego’s article cites Elsevier submissions rising from 2.7 million in 2022 to 4.2 million in 2025, a 56% increase. Reviewers generally volunteer their time. Zhang argues that AI adds another difficulty: it can make weak research look more polished, complicating the task of distinguishing strong papers from plausible-looking ones.
He also points to what he calls “shadow AI”: reviewers using AI tools independently, sometimes without disclosing that use. That leaves journals with little visibility into how the tools affect judgments. His proposal would make the AI assessment an explicit, shared part of the process rather than an undisclosed influence on an individual review.
Authors get a chance to challenge the AI
The author’s place in that sequence is deliberate. An AI assessment would not simply arrive on a human reviewer’s desk as an unanswered critique. Authors would first have an opportunity to correct errors, explain unclear points or address legitimate objections. Human reviewers would then see both the initial assessment and the response, alongside the paper.
Zhang says this could shift routine checks away from human reviewers, freeing them to concentrate on originality and significance—questions requiring scholarly judgment. He also argues that giving authors a voice, and making the AI assessment visible to everyone involved, could make the process more transparent and accountable.
Shared evidence can become a shared bias
The same visibility creates the proposal’s central risk. Reviewers who read one common AI assessment may become anchored to it: their later judgments could follow its framing rather than develop independently. Zhang warns that journals could receive variations on one opinion instead of several distinct assessments.
The paper examines that tradeoff through mathematical modeling. Its finding is conditional, not a demonstrated improvement across working journals. Under some conditions, the process could improve decision accuracy while reducing human review time. But excessive reliance on the AI could erase those benefits, even though the proposed workflow still includes human reviewers.
Zhang identifies other questions the model cannot settle: whether authors will learn to game the process and whether unconventional research could be overlooked. These are unresolved risks, not findings that either outcome has occurred. Reviewer dependence is also a practical question for experimentation, beyond the mathematical treatment.
A version is being tested at the University of Arizona’s Workshop on Responsible TecH, Zhang says. That experiment is intended to examine whether the process can ease review work without losing the independent judgment it is designed to support.
Reader comments
Newest comments first. Replies stay oldest first.