Surge in AI Content Overwhelms Academic Peer Review System, Effectiveness of Traditional Peer Review as Quality Control Declines
The volume of submissions to academic journals and the generation of review reports have surged sharply due to the large-scale involvement of generative artificial intelligence, leading to a decline in the effectiveness of traditional peer review as a quality control mechanism.
The journal "Accountability in Research" reported a 60% increase in submissions this year, with editor David Resnik noting that many submissions are suspected to be AI-generated, involving the fabrication of academic papers and public data mining. Additionally, editors need to send about 20 invitations to receive 2 valid review responses, with some responses also suspected to be AI-generated. A survey by Frontiers of 1,645 researchers revealed that 53% of peer reviewers have used AI tools to assist in reviews, up from about 29% the previous year, with 59% using it for writing review reports and 29% for summarizing manuscripts or checking references.
An analysis by "Organization Science" of 6,957 submissions and 10,389 review reports since 2021 found a 42% increase in submissions following the release of ChatGPT, while the number of manuscripts independently written by humans actually decreased. By early 2026, most manuscripts showed signs of AI involvement, with over 70% of AI content categories experiencing the fastest growth. In the International Conference on Learning Representations (ICLR) 2026, about 75,800 reviews showed that 21% were completely AI-generated, and more than half exhibited traces of AI assistance, with fully AI-generated reviews scoring higher and being longer on average.
Several publishing institutions have clear policies prohibiting the uploading of unpublished manuscripts to public AI tools to protect confidentiality. Springer Nature, Wiley, and Elsevier restrict reviewers to using AI only for language polishing and require disclosure. However, actual usage is widespread, and AI reviews perform well in correctness checks but provide almost no incremental information on originality and significance. Related studies indicate that rewriting manuscripts through prompt models can systematically improve AI review scores, creating a "paper laundering" effect; the similarity between AI reviews is significantly higher than that of human reviews, exhibiting a "hivemind" homogenization.
On average, editorial offices need to send out 4.5 invitations to obtain one completed review, doubling since 2018. AI-generated papers often overlook methodological flaws, continue to cite retracted literature, and prefer AI-style writing, resulting in superficially complete but substantively weak manuscripts passing initial screenings. Sakana AI's multi-layer review framework can capture 73% of core claim errors on synthetic error detection benchmarks, but only about 16% of true retracted papers are accurately matched.
In market mechanisms, authors use AI to mass-produce submissions under the "publish-or-perish" incentive to increase publication opportunities, while reviewers use AI to reduce time costs and invitation burdens. Journal editors face a mismatch between exponentially increasing submission volumes and insufficient review supply. Funding and attention are directed toward AI tool providers and detection service providers, with traditional publishing processes and human expert time becoming major pressure points; AI-assisted reviews may improve efficiency in some processes but simultaneously reduce review diversity and the ability to recognize innovative contributions.
Source: Public Information
ABAB AI Insight
The peer review system has long relied on a few senior scholars to bear a heavy load of reviews. From 2018 to 2025, the number of invitations needed to obtain one valid review has increased from about 2.25 to 4.5, while the volume of submissions to conferences and journals continues to grow exponentially without a corresponding expansion in the supply of qualified reviewers, creating a structural bottleneck.
AI tools are used by authors to accelerate manuscript production and by reviewers to draft reports, with resources effectively flowing to model service providers and detection companies; authors are motivated to lower publication barriers in response to career incentives, while reviewers aim to reduce the time spent on each review, and journals attempt to manage the increased throughput with limited editorial resources.
This is similar to the previous path of paper mills producing low-quality papers in bulk, only the production tools have shifted from human collaboration to model generation; the current phase is in a transition between expansion and control, where AI has increased output scale but has not yet formed an effective quality control alternative.
Essentially, this represents a combination of technological replacement and regulatory changes: models can replace some correctness checks and text generation but cannot substitute for judgments on the significance of contributions, while confidentiality rules and disclosure requirements lag behind actual usage, leading to a shift in quality control from human judgment to outputs that can be manipulated stylistically, thereby amplifying homogenization and weakening the identification of methodological flaws.
ABAB News · Cognitive Laws
Tools lower barriers while amplifying system load
Incentives determine output, quality determines filtering failure
Correctness can be automated, but contribution judgment still requires humans