Peer review is often described as if it were a single, settled practice: an editor sends a manuscript to two or three anonymous experts, they recommend acceptance or rejection, and the journal decides. In reality, the system has never been uniform, and it is now more varied than at any point in its history. Journals differ on who knows whose identity, whether reports are published, whether review happens before or after a paper becomes public, and whether a review can travel with a manuscript from one venue to another.
These changes are not driven by a single problem. Some respond to concerns about bias and accountability, some to the wasted effort of reviewing the same paper repeatedly for successive journals, some to the growth of preprints, and some to the long-standing complaint that reviewing is unrecognised scholarly labour. More recently, generative AI has forced publishers to set explicit rules for what reviewers may and may not do.
This explainer maps the main models, the infrastructure for recognising reviewers, and the policies now governing AI use, and it looks honestly at how little is known about which approaches work best. Pressures on reviewer supply and the question of whether software could take over the reviewer's role are covered in JCF Media's analyses of the peer review bottleneck and AI and human peer review.
Who knows whom: anonymity models
The traditional baseline in many sciences is single-anonymous review, in which reviewers know the authors' identities but authors do not know the reviewers'. Double-anonymous review hides both sides and is common in the humanities, social sciences and parts of medicine and engineering; its aim is to reduce bias linked to authors' reputation, institution, gender or nationality, though in small fields reviewers can often guess who wrote a paper.
At the other end sits open-identity review, in which reviewers are named to authors and sometimes to readers. Advocates argue that it encourages courteous, careful reports and makes conflicts of interest easier to spot. Critics worry that early-career researchers will hesitate to criticise senior colleagues.
Publishing the reports
A separate question is whether review content is published. Transparent peer review publishes reports and author responses alongside the article, usually keeping reviewers anonymous unless they opt to sign. Nature made this standard for new research submissions from June 2025, after offering it as an option since 2020; Nature Communications has applied it to all direct submissions since 2022. Many other journals offer it on an opt-in basis.
Published reports give readers insight into how a paper was scrutinised and turn reviewing into a visible, citable contribution. They also change the audience a reviewer writes for, which some journals address with guidance on tone and redaction.
Collaborative and consultative review
Some journals have moved away from wholly independent reports. In collaborative or consultative models, reviewers see each other's reports and discuss them, often with the editor, before a consolidated decision letter is sent. eLife used a consultative process for years before adopting its current model. The approach can resolve contradictory demands that would otherwise leave authors trying to satisfy incompatible requests, at the cost of more coordination time for editors and reviewers.
Review that travels: cascading and portable models
Cascading review keeps reviews within a publisher: if a manuscript is rejected from one journal, the authors can transfer it, with its reports, to another title in the same portfolio. It saves time but is limited to a single publisher's journals.
Portable, journal-independent review goes further. Review Commons, a project of ASAPbio and EMBO, organises one round of peer review in the life sciences before a manuscript is submitted to any journal. Authors receive a Refereed Preprint — the manuscript, the reports and their response — which they can send to one of its affiliate journals from publishers including EMBO Press, eLife, PLOS, The Company of Biologists and Rockefeller University Press, and which can be posted to bioRxiv.
Review of preprints
The fastest-growing area is review attached to preprints rather than to journal submissions.
- eLife, since 2023, publishes every manuscript it sends for review as a Reviewed Preprint with public reviews and an eLife assessment of the work's significance and strength of evidence. It no longer makes accept-or-reject decisions after review; authors decide when to declare a version of record.
- Peer Community In is a non-profit network of thematic communities in which recommenders manage review of preprints by at least two reviewers and, if the evaluation is positive, publicly recommend them. Authors can then publish in the diamond open access Peer Community Journal, submit to a journal that accepts PCI recommendations, or go elsewhere. The process is free to authors and readers.
- PREreview runs an open-source platform for reviewing preprints, with training, mentoring and live collaborative reviews intended to widen participation in peer review. Its reviews are deposited in a Zenodo community.
- MetaROR, launched in November 2024 with involvement from the Research on Research Institute, applies the publish–review–curate model to metaresearch: reviews and editorial assessments are public, and there is no binary accept-or-reject gate.
These services share a premise: that the scholarly value of review lies in the reports themselves, not only in a journal's yes-or-no verdict.
Structured review forms
Alongside these models, many journals now ask reviewers to answer specific questions — whether the conclusions follow from the data, whether statistics are appropriate, whether data and code are available — rather than or in addition to writing a free-form report. Structured forms aim to make reviews more consistent and complete, to make it easier for editors to compare reports, and to help less experienced reviewers know what is expected. They also produce data that journals can analyse.
Recognising reviewers
Reviewing has traditionally been invisible on a CV. Two infrastructures now address that. ORCID records can include peer review activity, added by trusted organisations such as publishers and recognition services rather than typed in by the reviewer. Clarivate's Web of Science Reviewer Recognition Service, formerly Publons, lets reviewers record and verify reviews — including anonymous reviews and reviews of manuscripts that were never published — and can export verified records to ORCID. Publishers using it typically display only the year and journal, keeping report content confidential unless journal and reviewer agree otherwise.
Recognition is not reward. Few institutions weigh review activity seriously in hiring or promotion, and verified records cannot by themselves show whether a review was any good.
AI-use policies for reviewers
Generative AI has prompted publishers to state explicitly what reviewers may do. Elsevier's policy tells reviewers not to upload manuscripts — or their own reports and correspondence — into generative AI tools, citing confidentiality, proprietary rights and data privacy, and says generative AI should not be used to assist in the scientific review of a paper, because critical assessment lies outside the technology's scope. Springer Nature similarly asks reviewers not to upload manuscripts into generative AI tools, holds them accountable for their reports, and requires them to declare any AI support used in evaluating a manuscript's claims.
The common thread is confidentiality and accountability: the reviewer, not the tool, is responsible for the report. Publishers are, at the same time, deploying their own controlled AI tools for editorial screening, which they distinguish from reviewers' use of public services.
What the evidence says
For all this experimentation, the evidence base is thin. A Cochrane systematic review by Jefferson and colleagues in 2007 found little empirical evidence to support editorial peer review as a means of ensuring the quality of biomedical research, while stressing that lack of evidence is not evidence of ineffectiveness and that peer review is methodologically difficult to study. A randomised trial at The BMJ, published in 1999, found that asking reviewers to reveal their identities to authors made no meaningful difference to review quality, recommendations or time taken — but significantly increased the likelihood that invited reviewers would decline.
Newer models, including preprint review and publish–review–curate, have not yet accumulated comparable trial evidence. Most claims for and against them rest on experience, surveys and logic rather than controlled comparison.
Conclusion
The next generation of peer review is not a single model replacing the old one. It is a widening menu: anonymity choices, published reports, collaborative consultation, portable reviews, preprint assessments, structured forms, recognition records and explicit AI rules, combined differently by different communities. What should guide editors and publishers through that menu is less fashion than fit — what their field needs, what their reviewers will accept and what can be evaluated. The most useful contribution journals could make now would be to treat their own changes as experiments, measure them and publish the results.
