Every editor and reviewer wants reviews that are thorough and timely, but the volume of submissions continues to grow, and the available human expertise is therefore limited. Artificial intelligence (AI) is now helping to bridge that gap, complementing human expertise rather than replacing it.

AI can already read thousands of papers, summarize their content, and check for inconsistencies in text. Several companies offer this service, including Consensus, Elicit, and Undermind. At ReviewerZero, we can also flag missing citations, potential plagiarism, or unusual statistical patterns. What used to take days or weeks of manual checking can now happen in minutes. This increase in speed and scale enables editorial teams to cover a broader range of topics and identify issues earlier in the process.

But speed alone is not the goal. The real opportunity lies in raising the quality of peer review. AI enables reviewers to focus more on substance, such as whether a paper’s reasoning is sound, its claims are proportionate, and its contribution is meaningful. When used well, AI can make peer review more comprehensive, fair, and constructive, not just faster.

We’ve been building agentic systems that can mimic the best aspects of human feedback and machine precision. Two recent projects, MAMORX and REMOR, explore how AI can write structured, evidence-based review reports. MAMORX acts as a multi-agent reviewer, examining the text, figures, references, and statistical patterns to summarize the strengths and weaknesses of a manuscript. REMOR is designed to identify methodological flaws and offer feedback.

Still, my work has made one thing clear: no matter how capable AI becomes, it cannot replace the human sense of judgment that defines good peer review.

Machines don’t understand (yet) the historical context of an idea or the subtle implications of a new result. They can measure coherence and consistency, but not originality or motivation. More importantly, they cannot take responsibility for their actions. In this balance, AI is envisioned as a tool that amplifies human capacity rather than substitutes for it.

Daniel Acuña, Founder of Reviewer Zero AI and Associate Professor at the University of Colorado at Boulder (on leave)

In 2024, I was invited by the DeSci Foundation to give a talk on the future of peer review and reproducible science, and it’s a pleasure to revisit and update those ideas here. Here’s what I believe the future of peer review will look like: AI tools will provide recommendations on how each aspect of a study could be improved, from planning, execution, and reporting. Researchers will also use AI tools to obtain templates for study protocols, data management and statistical analysis plans, as well as results reporting manuscripts. It is likely these tools will be able to even write these for researchers in their own “voice”. With the help of AI, researchers will be able to more easily produce fully computationally reproducible manuscripts that showcase the full data pipeline from the raw data to the reported results, tables, and figures.

While it might take some time until we see studies of AI cost-effectiveness, an ideal scenario, from a user perspective, would be one where an individual prompts one or several AI tools about anything in their current design or reporting that can be improved. If we reach that point, the main question before me is: will human review be needed before publication, or will preprints and AI-passed checks become the predominant way of scientific communication?

In today’s publishing world, almost the only way that researchers can be “persuaded” to follow the reporting guidelines fully, to share their code and data, or to create one-click reproducible manuscripts, is when a funder or a journal forces their compliance (I recommend checking the contributions from Robert Thibault and colleagues and from Simine Vazire for the recent Peer Review Congress on this topic).

No universities or research institutions have yet been bold enough, or had sufficient resources, to require and monitor the same for all their employees. The main arguments for why researchers are not doing so seem to be the lack of time, ease-of-use tools, and job security (i.e., pressure to publish) needed to conduct science at the highest possible standards. The hope is that AI will make it easier to follow and verify those standards more efficiently.

Mario Malicki, Associate Director of the Stanford Program on Research Rigor and Reproducibility (SPORR)

Reviewers just want to have fun. There is a looming crisis in peer review. This is, in part, a crisis in value: a growing number of researchers either feel like peer review is not valuable enough to be worth doing, not valuable enough to be worth paying for as a service from a journal, or both.

This is a problem because peer review is a key mechanism of self-correction in science and one of the few tangible services that publishers can point at when justifying article processing charges. Without it, science will fall over and publishers won’t make money.

At Evidence-Based Toxicology (EBT), we think a lot about peer review. As the journal of a voluntary collaboration for advancing best practices in toxicology, we are a test bed for ideas about how scientific publishing can more positively contribute to the healthy functioning of the research ecosystem. At the same time, we are a journal from a commercial publisher and are expected to turn a profit. What level of service can we afford to provide, in proportion to what people are willing to pay?

The foundational principle of our approach at EBT is that peer review should have intrinsic value for the reviewer: they should be doing it for themselves, not out of a sense of duty. (This is not to deny that duty is a powerful motivator, but it should not be your leading offer.)

For peer review to have intrinsic value, a manuscript needs to be enjoyable to read and intellectually stimulating to engage with.

However, reviewing a paper that is incomplete and difficult to understand is not fun. The handling editor’s job at EBT is to make sure a manuscript is sufficiently complete, coherent, and correct that in nearly all cases a reviewer’s only task will be to enjoy improving it.

“Improve” is the key. At EBT we do not reject manuscripts (except when they fail integrity checks) because rejection after peer review does not help the authors and was not worthwhile for the peer reviewers. Instead of rejecting, we explain what needs to be done to get the manuscript ready for peer review — and only advance a manuscript to review once the editor is confident that the peer-reviewers will enjoy commenting on it.

EBT’s high level of editorial intervention necessitates collaborative review. We offer authors calls with the handling editor. We involve the reviewers if they consent. Collaboration builds networks, and networks are important to researchers, which further reinforces the intrinsic value of the peer review process.

The final ingredient is transparency. At EBT, we require a submitted manuscript to be a registered preprint before we will consider evaluating it. We also publish our manuscript evaluations. This makes us accountable for our decisions, surfaces for the reader useful points of scientific discussion, and allows reviewers to be credited for their work.

We think peer review has to move in this direction. Publishing is expensive and it has to be worthwhile — not only for the authors who pay directly, but the reviewers who pay in kind.

Paul Whaley, Editor-in-Chief of Evidence-Based Toxicology

What does the future of peer review look like through different lenses? In his piece on the future of peer review, Daniel Acuña of Reviewer Zero AI makes the claim that AI can help address the problem of the mismatch between ever-rising submissions and the available human expertise. One of the strengths of AI is that it can read and summarize thousands of papers, checking for textual inconsistencies as it goes along. This technology is being pioneered by several companies, including Daniel Acuña’s Reviewer Zero, which can also run an integrity check for issues such as plagiarism, missing citations, and unusual statistical patterns. This is much faster than human efforts and frees up editorial resources. However, he also points out that speed alone is not the goal, but rather to raise the quality of peer review. He proposes that by handling the more mundane parts of the evaluation, AI can free up reviewers to focus more on substantial issues. Daniel Acuña also highlights how AI can generate structured, evidence-based review reports. He discusses two tools that together can analyze all aspects of a submitted manuscript.

Acuña sounds a note of caution that no matter how powerful AI becomes, it cannot replace human judgment. He names a number of human qualities AI cannot yet emulate, such as knowledge of context, a lack of understanding of originality, and that it cannot take responsibility as is expected for an author. Acuña sees AI as a tool to amplify human capacity rather than replacing it.

For his part, Mario Malicki, Associate Director of the Stanford program Research Rigor and Reproducibility (SPORR), discusses the future of peer review through the lens of compliance. He also believes AI can be of help here, but his focus is not so much generative as Acuña’s, but rather about helping authors and journals to uphold integrity standards by, for instance, assisting with compliance checks and screening manuscripts that fall short in this.

Paul Whaley, who is the Editor-in-Chief of Evidence-Based Toxicology, homes in on another aspect of the crisis in peer review, which is motivating reviewers, asking the question about value for the reviewer. Whaley proposes that the motivation for the reviewer should not be a sense of duty to a community or any kind of implicit quid pro quo. Rather, he proposes that reviewing becomes enjoyable, or at least less painful, if the manuscript is in good shape before it is sent to reviewers. He draws on the example of his journal to make his point. His journal’s collaborative review and pre-review manuscript improvement by handling editors results in a different and hopefully more satisfying reviewer experience.

All the contributing pieces more or less base their thinking on the idea that the present system is not going to be enough going forward, and perhaps is already failing the research enterprise. This seems to be a commonly held view in scholarly publishing, existing as a backdrop and existential threat in the same way that Moore’s law looms over computing.

In all the contributions, there is a recurring theme of improving the reviewer experience by not subjecting them to tedious grunt work, unpacking the reviewing endeavor into more or less intellectually demanding parts, where the most cognitively challenging tasks are reserved for and require the human intellect with all its expertise and power of judgment. However, the matter of what constitutes tedious grunt work is itself a matter of judgment. For Whaley’s example, by shifting so much of the work of getting a manuscript to a reviewable state to editors, it runs the risk of short-circuiting reviewer input and not getting the independent perspective that is needed.

In the case of both Acuña and Malicki, the degree of AI involvement represents another judgment point regarding how one categorizes work that needs a reviewer’s personal attention and work that can be outsourced to AI. It may not always be such an easy task to figure out where the line should be drawn. Similarly, for assistance with compliance, setting the guidelines will still require human input and oversight to ensure that we leave the key decisions in the hands of humans.

It is possible that the task of reviewing itself requires an update. Shifts have been made in the standard mode of peer review at several junctures in the history of scholarly publishing. Journals started involving external reviewers rather than letting editors make decisions by fiat. Anonymity, for reviewers and also for authors, was deployed as a strategy in the aim of achieving a more objective process. Editors started consulting several reviewers in order to obtain a multitude of perspectives.

In his book, The Scientific Journal, Alex Csiszar makes the claim that changes to peer review occur at times of societal upheaval, correlating with crises in the public standing of science. With the rise of AI and its attendant flood of AI-generated content masquerading as real research, not to mention the current politicization of science, academic publishing and peer review are at a crossroads. We need validation, and that validation has to develop in a sustainable fashion to prevent burnout, preserve engagement, and above all, keep the peer in peer review.

Daniel Ucko, Head of Ethics and Research Integrity, American Physical Society