Welcome to peer review: AI writes, AI judges, humans applaud

Officially, peer review is the gold standard of science. In practice, it is a volunteer run bottleneck trying to cope with an industrial scale flood of manuscripts.

Annual outputs indexed by Crossref have grown more than fourfold since 2004, surpassing 10 million records by 2024. On the ScholarOne platform, journals received 33% more submissions in the first quarter of 2026 than in the same period in 2025, with the year on year growth rate itself more than doubling.

At the International Conference on Learning Representations (ICLR), one of the flagship AI conferences, submissions jumped from 1,013 in 2018 to 19,814 in 2026, a tenfold increase in under a decade.

Much of this expansion is concentrated. Analyses suggest that over 70% of the net increase in yearly published articles is driven by a small cohort of high volume commercial publishers, with MDPI – Multidisciplinary Digital Publishing Institute – a major contributor. Meanwhile, peer review still depends on unpaid academics who are increasingly saying ‘no’.

Reviewer invitation acceptance rates have fallen sharply: from around 56% in 2020 to roughly 36% to 39% by 2024-25 in documented cases. Editors report needing four to five invitations on average to secure one reviewer, with some journals requiring over 30 requests to obtain two reviews. Median time from submission to acceptance has lengthened by several weeks in cross field analyses.

The burden is also uneven. IOP Publishing – the Institute of Physics Publishing – data show that 30% of reviewers in high income countries report being overloaded, compared with only 10% in low and middle income countries, where researchers face biased selection, weak incentives and limited access to verification infrastructure.

It is no surprise that, in a global poll of 2,115 stakeholders, 40% chose “Peer Review Capacity: Volume, speed, and quality” as the theme for Peer Review Week 2026.

The polite way to describe this is ‘structural strain’. The honest description is that the system is quietly breaking.

Enter generative AI: More papers, fewer brains

Just as the reviewer pool started to shrink, generative AI arrived to make everything faster, longer and less readable.

Studies of top management and AI journals show that large language model (LLM) assistance is associated with a 42% increase in manuscript submissions since late 2022. These AI heavy papers score worse on readability metrics, use denser jargon and face roughly double the desk rejection rate compared with more human written submissions.

At the journal Organization Science, the fastest growing category post ChatGPT was papers scored at 70% or higher AI content; these are significantly less likely to receive a revise and resubmit decision.

AI is not only writing more papers; it is also reviewing them.

At ICLR 2026, an analysis of 75,800 peer review reports found that 21% were fully AI generated, and over 50% showed some form of AI involvement.

AI generated reviews tend to be longer (sometimes up to 3,000 words), heavily formatted and packed with dozens of questions and weaknesses, yet they exhibit lower information density and systematically higher scores (average 4.43 versus 4.13 for human reviews) despite correlating with lower overall paper quality.

The downstream effect is visible in error rates. A systematic analysis of published AI papers found that the average number of objective errors per paper rose from 3.8 at NeurIPS 2021 to 5.9 at NeurIPS 2025, a 55.3% increase.

NeurIPS – the Conference on Neural Information Processing Systems – is one of the premier venues for machine learning research. When multiple AI models are used across writing, reviewing and meta reviewing, there is a real risk of what some researchers now call ‘synthetic consensus cascades’: correlated model biases eliminate statistical independence in peer review, making systemic errors harder to detect.

In short, we have built a system where AI writes more papers, AI reviews more papers, and humans are left to pretend they are still in control.

Voices from academic leadership and the push for policy reform

Senior academic leaders and funders increasingly argue that solving this crisis requires decoupling evaluation from raw volume and journal metrics.

Marc Schiltz, president of Science Europe, has warned that the entrenched ‘publish or perish’ culture and reliance on quantitative bibliometrics have driven scholarly communication into a “blind alley”, urging institutions to transition to qualitative, narrative based evaluations.

Michael Murphy, president of the European University Association (EUA), has similarly called on universities, researchers and funding agencies to actively change how research is assessed to halt the degradation of the scientific record.

Writing in 2025, Esma enel, an ambassador for the Asian Council of Science Editors (ACSE), recalled her video contribution to the "Voices from ACSE" series for Peer Review Week 2025, emphasised that “AI belongs in supporting tasks such as screening and linguistic clarity, but it does not belong in interpretive judgment or the ethical reasoning that peer review fundamentally requires”.

Metascientists such as Dr Michael Bruno and Ludo Waltman have argued that AI alone cannot rescue peer review without a “revolutionary change” to academia’s underlying incentive structures; research assessment reform and publishing reform must proceed together.

These perspectives are now being codified in policy:

• The Stockholm Charter on Science Publishing (2026) mandates ultimate human control over peer review and prohibits fully automated AI evaluations or decision making in editorial processes.

• CoARA (Coalition for Advancing Research Assessment), ARRA (Agreement on Reforming Research Assessment) and DORA (San Francisco Declaration on Research Assessment) commit signatories to eliminate inappropriate uses of Journal Impact Factors and h index in research assessment, pushing towards qualitative, portfolio based evaluation.

• Australia’s NHMRC (National Health and Medical Research Council) Action Plan for Responsible Research Assessment (2025-2030) introduces measures such as a ‘Top 10 in 10’ publication cap per assessment cycle, application centric reviewer matching, and the Research Impact Track Record Assessment (RITRA) framework to shift focus from volume to verified impact.

The mechanics of verification capacity and incentive collapse



Random audits, trust markers and other inconvenient fixes

If the problem is structural, the solution must be structural. Reformers are proposing mechanisms that make verification cheaper, more visible and harder to game.

One proposal is Random Instrumented Audits: journals commit a randomised sample of accepted manuscripts to independent validation teams pooled across multiple organisations.

Audits would re run analysis scripts on raw data, verify mathematical proofs, check baseline configurations and hyperparameter seeds, and publish outcomesas additive, non punitive addenda. This introduces negative expected utility for fragile claims and makes truth seeking rational again.

Parallel ideas include:

• Publish Review Curate (PRC) architectures, which decouple immediate preprint dissemination from ongoing, public evaluation. The PRC Alliance, formed by COAR (Confederation of Open Access Repositories) and ASAPbio (Accelerating Science and Publication in Biology), along with native OJS 3.6 (Open Journal Systems version 3.6) workflows, support this model.

• NISO Trust Markers, approved in May 2026 by NISO (the National Information Standards Organization), which aim to establish a tiered publication level metadata framework tracking data/code availability, preregistration, open review reports and documented author responses.

Together, these mechanisms aim to shift incentives from ‘publish more’ to ‘publish verifiably’.

Verification first AI: audits before scores

A new generation of tools is being depoyed as adversarial auditors that generate inspectable evidence trails before human reviewer bandwidth is consumed.

In a typical workflow:

Stage 1: Automated Auditability Screening

• Correctness Check tools extract core claims and run non trivial error checks (invalid formulas/logic), evidence sufficiency checks (baseline gaps) and novelty checks against literature, producing a reader friendly report that routes human attention to high risk passages.

• Reference Validity Check tools resolve each bibliography entry against canonical registries (Crossref, arXiv) and assign confidence scores. In evaluations on corrupted and hallucinated reference strings, such tools have achieved around a 93% detection rate of problematic citations.

Stage 2: Targeted Human Verification and Public Audit

• Focused reviewer bandwidth on core claims flagged by Stage 1.

• Random post acceptance reruns and public addenda.

Tools such as CSPaper (part of Scholar7) and SciCoQA (Scientific Paper-Code Quality Assurance) operationalise this vision. CSPaper, selected by TMLR (Transactions on Machine Learning Research) for an official AI reviewer pilot, has shown high agreement with action editors on accepted papers and zero false passes on papers rejected on soundness criteria in independent evaluations.

SciCoQA frames verification as paper-code alignment auditing, turning claims into measurable, reusable QA artifacts; current LLMs detect only about 45% to 46% of real discrepancies, underscoring both the difficulty of the task and the need for better tooling.

The emerging lesson is that AI is most credible for verification when it produces inspectable evidence trails, not opaque scores.

The uncomfortable conclusion

Peer review is not fine. It is overstretched, increasingly automated and drifting towards proxy driven evaluation.

The choice is not between ‘more papers’ and ‘fewer papers’. It is between a system that rewards volume and a system that rewards verified knowledge.

Random audits, trust markers, PRC models and verification first AI are not silver bullets, but they are concrete steps towards a system where ‘peer review’ once again means something close to what the phrase promises.

If we do not make that shift, the safest prediction is not a sudden collapse but a slow, quiet transition to a world where most papers are read by machines, reviewed by machines and cited by machines – while humans continue to pretend they are in charge.

Dr Hemachandran Kannan is vice dean in the School of Business and director of the AI Research Centre at Woxsen University in Hyderabad, India. Dr Raul Villamarin Rodriguez is vice president at Woxsen University, India.

Disclaimer: The views expressed are those of the authors and do not necessarily reflect the official policy or position of Woxsen University or its partners.

This article is a commentary. Commentary articles are the opinion of the authors and do not necessarily reflect the views of University World News.