Human Judgement Is Indispensable

In 2020, I began serving as co-editor-in-chief of the international journal Applied Geochemistry, later becoming editor-in-chief and completing my term this year. Over those six years, artificial intelligence moved rapidly from supporting tasks such as language editing and literature searches into manuscript writing, peer review and the dissemination of research findings. Scholarly publishing has changed profoundly in the process.

In the day-to-day running of a journal, the handling associate editor usually organises peer review, with experts in the relevant field providing their professional assessments. The editor-in-chief's primary responsibility is to establish the journal's scholarly direction and quality standards: the questions researchers are encouraged to ask, the evidence needed to support their conclusions, and the approach to disagreement and uncertainty. This role complements the judgements made by editors and reviewers. As AI enters scholarly publishing, editors-in-chief must also define appropriate boundaries for its use and uphold the professional integrity and credibility of peer review.

Over the years, I have become increasingly aware of how difficult it is to find suitable reviewers. Every discipline has only a handful of journals at its core. Even these journals often need several rounds of invitations to secure appropriate experts; the many other specialist journals face this difficulty even more widely. Submission numbers can rise rapidly, but the time and attention that experts can devote to them cannot grow at the same rate.

Peer review underpins the quality of research and trust within the scientific community. Yet it has long received less recognition and reward than its value warrants. Publications, grants and citations have a clear place on a CV. The time and judgement required to review a manuscript carefully are much harder to make visible. A reviewer may recalculate data to check a figure or read extensively to establish whether a conclusion holds. Often, the only acknowledgement is an automated thank-you email.

Everyone wants their own papers to receive careful review. Under current assessment systems, however, researchers often have to give priority to outputs that count towards their evaluations. Over time, the most conscientious reviewers face the greatest opportunity costs, making them more likely to reduce their reviewing commitments or withdraw altogether.

AI has intensified this tension while making it harder to see, although the shortage of reviewers has deeper roots. It can rapidly produce a well-structured, carefully worded report. Understanding a research question, checking data and weighing evidence still require substantive work. Editors must now assess whether the reports they receive reflect real, sufficiently thorough reading and judgement. A report that appears complete can conceal a review that has never truly taken place.

I sometimes find myself contemplating an unsettling possibility: for a substantial share of future papers, the first "reader" to process the entire text may be AI. An author might use AI to write a paper, a reviewer might use it to generate a report, and other researchers might later learn of its conclusions mainly through AI-generated summaries and recommendations. A paper could then pass through the entire process of publication and dissemination without any person ever reading it carefully from beginning to end.

AI can be useful as a paper's first reader, helping researchers search, screen and organise an overwhelming volume of literature. The danger arises when people stop returning to the original text at critical points, checking the data behind the conclusions, or asking whether the claims extend beyond the evidence. Machines can process papers. The scientific community remains responsible for understanding them, judging their merits and bearing responsibility for the research.

In this setting, peer review becomes even more important. It may be the most consequential complete reading a paper receives before publication. The value of a review lies in identifying the issues that determine whether a study's claims hold and providing verifiable grounds for that assessment. The number of comments is a poor measure of that value. Tools can help improve the language, while reviewers and editors must make the key scientific judgements and take responsibility for them.

The governance of AI calls for more than blanket bans or suspicion of researchers, particularly young researchers who are keen to use it. Using tools to improve language or organise material is compatible with research integrity. What matters is whether researchers retain control over their own reasoning, verify generated content, can explain the research process and accept responsibility for the final work. Assessment systems must also be consistent: they should not pressure early-career researchers to publish more and faster while treating all reasonable use of tools as an ethical failing.

High-quality reviewing should receive recognition commensurate with its value. Universities, research institutions and journals can include verified peer-review contributions in assessments of academic service and take them into account in editorial-board appointments and career progression. Such recognition must go beyond counting reviews. It should reward the identification of critical issues, the provision of reliable evidence, and contributions that help authors improve their research and editors reach sound decisions. Journals should also reduce ineffective and duplicated review, set reasonable deadlines, and make their requirements for confidentiality, disclosure and verification clear.

Looking back, I value the process in which authors refine their arguments, reviewers read page by page, and both engage in sustained discussion of the evidence. Authors and reviewers may never have met, yet a rigorous review and a well-reasoned response can create a sense of fellowship through scholarly exchange. The time saved by AI should support deeper reading, more rigorous verification and fuller discussion, rather than simply producing more text that no one reads carefully.

AI's entry into scholarly publishing is irreversible. Journals must preserve a clear, traceable chain of responsibility as working practices change: authors should be able to explain their central claims, reviewers should be able to justify their assessments, editors should be accountable for their decisions, and readers should be able to trace conclusions to their sources.

AI can contribute at every stage of research. At the critical points where the validity and value of knowledge are judged, human judgement remains indispensable.