Beyond Faster Peer Review: Building Stronger Evidence for Editorial Decisions

A faster peer-review process is not necessarily a better one. We can shorten turnaround times and generate more review text without improving the decisions that shape the scholarly record. For me, that is the central challenge in applying AI to peer review: are we accelerating the existing process, or strengthening the evidence behind editorial judgment?

Peer review is an essential layer of scholarly trust, not a certification that research is correct or reproducible. Its limitations should prompt us to rethink where assurance happens, what reviewers should contribute, and how responsibility continues after publication.

Protect the judgment we cannot automate 

The conversation about reviewer capacity often starts with recruiting more reviewers. I believe we should first ask how much avoidable work reaches those we already have. Missing reporting information, incomplete references, unavailable data, and unresolved technical issues can distract from the assessment that requires disciplinary expertise. Increasing capacity should mean protecting attention, not simply increasing the number of completed reviews. 

This is why quality assurance needs to move upstream. AI-assisted checks could help authors address problems during manuscript preparation and support editorial teams during submission and triage. Reviewers could then devote more attention to methodology, interpretation, originality, and significance. The objective should not be to help experts produce review text faster, but to help them review more deeply. 

From a product perspective, this argues for an adaptive workflow rather than a standard checklist applied indiscriminately. A clinical trial, a computational study, and a humanities article should not trigger identical checks. The system should surface evidence relevant to the manuscript and make clear which questions remain unresolved. 

Moving checks earlier only helps if their outputs are useful. Otherwise, we risk replacing reviewer fatigue with editorial alert fatigue.

Make responsibility part of the design

As I discussed in my earlier article on AI in peer review, AI speed must remain paired with human judgment. The practical question is how that principle shapes the evidence editors receive and the authority they have to act on it.

Putting that principle into practice requires more than a human approval button. Editors need to understand what a flag means, examine the supporting evidence, challenge an output, and know when to escalate. As AI systems gain autonomy, governance should become stronger and more visible.

For reviewers, LOCAD (Limitations, Ownership, Confidentiality, Accuracy, and Disclosure) remains a practical reminder of the responsibilities that accompany AI assistance.

An automated signal should initiate scrutiny, not substitute for a verdict. Accountability must remain identifiable throughout the workflow.

Move from checking manuscripts to testing research

The longer-term opportunity is to examine more of the research behind the manuscript. Checking whether a data-availability statement exists is useful. Establishing whether the underlying analysis supports the reported results is a different level of assurance.

I see an emerging direction toward “executable peer review.” For suitable computational research, AI agents could help reconstruct software environments, run code, compare outputs, and support checks of important assumptions identified by reviewers. These activities could give reviewers evidence that would otherwise be difficult or time-consuming to obtain. This is a direction to develop, not a capability to assume across every discipline.

Successful execution would not remove the need for scientific judgment or establish that the underlying research is valid. Reviewers would still need to assess whether the methods are appropriate, the assumptions defensible, and the conclusions warranted. More machine-generated evidence should enable deeper human assessment, not create another automatic approval signal.

Expansion should depend on whether the evidence helps editors make better-supported decisions at a proportionate cost. A ten-minute review is not progress if it produces a weaker decision.

For KGL and the publishers and societies we support, this is a content-lifecycle question, not just an editorial technology question. Evidence and responsibility should remain connected across submission, review, production, digital delivery, and post-publication correction. Publication should not mark the end of assurance.

The opportunity is therefore larger than automating yesterday’s workflow. It is to build a human–AI trust architecture that protects scarce expertise, makes decisions more accountable, and remains open to correction.

The question worth asking is not simply, “How quickly can we complete peer review?” It is, “How much more confidence can the process earn?”

Dr. Hong Zhou is Vice President, Product Management at KGL, where he leads product strategy and AI innovation for solutions serving researchers, librarians, and publishers. He can be reached at info@kwglobal.com.