AI in Peer Review: Speed Is Useful, but Trust Is Non-Negotiable

by Hong Zhou, VP of Product & AI Innovation

I was recently invited by the Asian Council of Science Editors to join an "Ask the Experts" podcast on AI and peer review. The questions were timely and practical: Can reviewers use AI? Can journals detect AI-generated text? Should authors verify AI-generated references? And what should journals do as AI becomes part of everyday scholarly publishing? 

My answer to all of these questions comes back to one principle: 

AI speed, human judgement. 

AI can help us read faster, check more, summarize better, and automate parts of the publishing workflow. But peer review and scholarly publishing are built on trust. That trust still depends on people, policy, governance, and accountability. 

Below are the key points I shared.

1. Reviewers can use AI, but confidentiality comes first

AI can be useful for reviewers. It can help summarize a manuscript, structure notes, suggest questions, or improve the clarity of a review. But we need to remember that a submitted manuscript is unpublished work. It may contain new ideas, confidential data, commercial value, or sensitive information. 

So the first rule is simple: 

Reviewers should not copy and paste confidential manuscripts into public AI tools unless the journal explicitly allows it and the tool has appropriate safeguards. 

A practical way to think about this is the LOCAD framework: 

L — Limitations
Reviewers should understand what AI can and cannot do. AI can summarize and organize information, but it can miss nuance, misunderstand methods, or generate inaccurate comments. 

O — Ownership
The manuscript belongs to the authors. Reviewers are temporary stewards of that work, not owners of the content. 

C — Confidentiality
Confidential manuscripts and peer review materials should only be used with journal-approved or institution-approved tools where data is not retained, reused, or used for model training. 

A — Accuracy
AI-generated summaries or comments must be checked carefully against the manuscript. Reviewers should not assume the AI output is correct. 

D — Disclosure
Where journal policy requires disclosure of AI use, reviewers should disclose it clearly. 

The integrity point is important: AI may assist the reviewer, but it must not become the reviewer. The human expert remains fully responsible for the review, the judgement, and the final recommendation.

2. AI-text detection should be treated as a signal, not a verdict

Many journals are understandably concerned about AI-generated text. But we need to be careful with the word "detect." 

In practice, it is becoming increasingly difficult to reliably distinguish between text that is fully AI-generated, AI-polished, translated, language-edited, or simply written in a very polished style. And in many cases, that distinction may not be the most useful question. 

The better questions are: 

Is the work accurate?
Are the references real?
Are the data valid?
Are the conclusions supported?
Has confidential material been protected?
Has AI use been disclosed according to journal policy? 

Current AI-text detection tools usually rely on signals such as writing patterns, predictability, word choice, or model-specific signatures. These can be useful, but they are not definitive. They can generate false positives, especially for non-native English writers, heavily edited text, or translated manuscripts. They can also miss AI-generated content that has been revised. 

So journals should not use AI detectors as "verdict machines." 

A more reliable approach is multi-signal integrity assessment. Instead of asking only "Was this written by AI?", journals should check for broader warning signs, such as fabricated references, irrelevant citations, citation-content mismatch, image manipulation, paper-mill patterns, suspicious reviewer identities, incoherent content, unsupported conclusions, or data concerns. 

If concerns remain, the best next step is usually a neutral author query, not an accusation. 

The goal should not be to police every use of AI. The goal should be to protect the scholarly record through transparency, accountability, consistent editorial standards, and human judgement.

3. Authors should verify every reference, especially when AI is involved

Authors are increasingly using AI tools to generate literature reviews, identify themes, and build reference lists. This can be helpful, but it also creates risk. 

AI tools can hallucinate references, invent DOIs, mix up authors and titles, cite papers that do not support the claim, or over-represent literature that is easier for the tool to access. In scholarly publishing, this is not a small formatting issue. References are part of the evidence base of the article. 

That means journals should require authors to take responsibility for every reference, no matter how it was found. 

Authors should confirm that: 

Each cited work exists.
The bibliographic details are correct.
The source has been read or checked.
The cited source supports the statement being made.
Any AI-assisted literature work is disclosed where required. 

This is a good example of responsible AI use. We do not need to ban AI-assisted literature work. But we do need clear accountability. 

AI can assist discovery, but authors must own the scholarship. 

Journals can support this through simple submission checks, including author confirmation of reference authenticity, AI-use disclosure, DOI validation, retraction checks, and tools that flag fabricated citations or citation-content mismatch.

4. Journals need AI governance, not just AI tools

AI should not be treated as a bolt-on feature or a magic solution. It needs to be built into publishing workflows carefully, with governance from the start. 

AI can support many parts of the journal workflow: technical checks, scope screening, reviewer discovery, reference checking, image analysis, summarizing reviewer reports, and workflow automation. But final editorial decisions, research integrity decisions, and ethical judgements must remain human-led. 

For journals, the best practices are clear. 

First, create clear AI policies. Authors, reviewers, editors, and staff need to know what is allowed, what is prohibited, and what must be disclosed. 

Second, protect confidentiality. Public AI tools should not be used for unpublished manuscripts or confidential peer review material unless the journal has approved safeguards. 

Third, use a risk-based approach. A grammar tool is relatively low risk. An AI-generated peer review, editorial recommendation, or research integrity escalation is much higher risk and needs stronger oversight. 

Fourth, train reviewers, editors, and staff. Many AI problems come not from bad intent, but from misunderstanding. People need practical AI literacy: when AI helps, when it misleads, and how to verify outputs. 

Fifth, monitor and audit AI tools over time. AI systems change. Their accuracy, bias, and behavior can drift. Vendor agreements should address data use, retention, model updates, audit rights, and incident response. 

The key question should not simply be, "Was AI used?" 

The better question is: 

Was the work honest, accurate, transparent, confidential, and accountable? 

Final thought: AI will change publishing, but trust will still define it 

AI will become a normal part of scholarly publishing. It will help publishers, journals, editors, reviewers, and authors work faster and smarter. But when AI capability moves faster than governance, risk increases. 

That is true not only for AI, but for every major emerging technology. 

For KGL, this is why responsible AI matters. The opportunity is not just to automate tasks. The opportunity is to help the scholarly publishing community build better, safer, more transparent workflows that protect the integrity of research. 

AI can accelerate the work. 

But trust still depends on people. 

KnowledgeWorks Global Ltd. (KGL) is the industry leader in editorial, production, eLearning, online hosting, and transformative services for every stage of the content lifecycle. We are your source for learning solutions, intelligent automation, research integrity, digital delivery, and more. Email us at info@kwglobal.com.