When people discuss artificial intelligence in scientific publishing, the concern usually centres on whether researchers are using chatbots to write manuscripts. That concern is understandable, but it may not be the most serious threat facing the scientific record. A more subtle problem is spreading through the literature: references that look entirely legitimate but point to papers that never existed. These fabricated citations are increasingly appearing in published research, often escaping both editorial checks and peer review. Rather than simply producing awkward prose, artificial intelligence is beginning to compromise one of science’s most fundamental institutions, its citation system.
The scale of the problem has become difficult to ignore. An audit of 2.5 million biomedical papers published between January 2023 and February 2026 identified more than 4,000 fabricated references embedded within 2,810 published articles. Even more concerning, the rate of fabricated citations increased more than twelvefold during this relatively short period, closely following the widespread adoption of large language models. When the analysis closed, almost all affected papers had received no correction, expression of concern, or retraction from publishers. The overwhelming majority remained part of the published scientific record.
Citation fabrication does not exist in isolation. It operates alongside a more established integrity threat: paper mills. These commercial operations produce and sell manuscripts, including authorship slots, to researchers facing intense pressure to publish. Investigations published in recent years suggest that hundreds of thousands of papers may contain characteristics associated with paper mill production, although only a small fraction have been formally corrected or retracted. Together, paper mills and AI-assisted citation fabrication represent two distinct but increasingly complementary mechanisms for introducing unreliable evidence into the scientific literature.
Indonesia occupies a particularly vulnerable position within this global landscape, not because its researchers are inherently more prone to misconduct, but because its research evaluation system has created unusually strong incentives to maximize publication output. Over the past decade, Indonesia moved from among the lowest-output research systems in Southeast Asia to one of the region’s most prolific producers of indexed scientific papers. This transformation followed the introduction of the Science and Technology Index (SINTA) in 2016 and subsequent regulations linking academic promotion and professional incentives to internationally indexed publications.
These reforms achieved their immediate objective. Publication numbers rose rapidly. Universities expanded international collaborations, faculty members became more active in submitting manuscripts, and Indonesia climbed regional publication rankings. When measured solely by bibliometric indicators, the reforms appeared successful.
Publication volume, however, does not equal research quality. Available analyses suggest that a substantial proportion of Indonesia’s indexed output came from conference proceedings rather than peer-reviewed journals. Conference proceedings have legitimate scholarly functions, particularly for disseminating preliminary findings, but they generally undergo less rigorous editorial and peer review than established scientific journals. When institutional reward systems value publication counts more heavily than publication quality, researchers naturally orient their behaviour toward whichever outputs are rewarded most reliably.
This distinction matters because incentives shape behaviour throughout scientific systems. Researchers rarely operate in isolation from institutional expectations. Promotion, financial incentives, institutional rankings, and research funding frequently depend upon measurable publication metrics. Whenever these indicators become the dominant criteria for professional advancement, demand emerges for services capable of accelerating publication. Paper mills exist because markets respond to demand. They exploit structural incentives rather than creating them.
The Indonesian Case
The incentive misalignments described above are already visible within Indonesian research. A recent review of retracted publications involving Indonesian-affiliated authors identified 148 formally retracted papers between 2015 and 2024. Retractions increased sharply after 2015 and reached their highest level in 2024. Medicine, engineering, and the social sciences accounted for much of this increase. Plagiarism remained the most common reason for retraction, but data fabrication, compromised peer review, authorship disputes, and publication in predatory journals also contributed substantially. Many affected articles appeared in APC-driven open access journals. The concentration there is not coincidental: rapid publication models reduce the friction that might otherwise deter submission of compromised manuscripts.
Artificial intelligence introduces a different challenge. Unlike paper mills, which manufacture entire manuscripts, large language models can fabricate references while producing otherwise plausible scientific writing. These references frequently combine authentic author names, plausible article titles, publication years, and journal names into citations that appear entirely routine. To reviewers scanning long reference lists, they look ordinary. Unless every citation is independently verified, fabricated references may remain undetected throughout peer review and publication.
The implications extend well beyond academic publishing. Review articles and systematic reviews occupy the apex of most clinical evidence hierarchies, including GRADE and OCEBM frameworks, because they synthesise findings from multiple primary studies to inform clinical practice and public policy. If fabricated references are embedded within these reviews, the evidence base those reviews produce becomes, in part, fictitious. According to the biomedical audit cited above, review articles exhibited substantially higher rates of fabricated citations than other publication types. For health research, where systematic reviews underpin clinical guidelines, this is more than a bibliographic problem. It potentially affects decisions about patient care and health policy.
From Individual Misconduct to Institutional Incentives
The instinctive reading of these developments is individual misconduct. Certainly, some researchers deliberately engage in unethical behaviour. Yet focusing exclusively on individual responsibility risks overlooking the institutional dynamics that make such practices increasingly attractive. Systems that reward publication quantity while neglecting quality, reproducibility, and editorial integrity reliably create conditions for exploitation. Paper mills and AI-generated citation fabrication are symptoms of deeper governance problems, not isolated ethical failures.
Structural problems require structural solutions. Research evaluation frameworks need to evolve alongside technological change, not merely penalize its misuse. Publication counts should carry less weight than indicators reflecting scientific influence, methodological robustness, and reproducibility. Universities should integrate automated reference verification into editorial workflows before manuscripts are submitted. National research integrity systems should classify fabricated citations as a distinct category of misconduct, so that prevalence can be tracked systematically rather than absorbed into the broader rubric of publication ethics.
Artificial intelligence did not create the incentive structures that encourage questionable publishing practices. Paper mills did not invent the demand for rapid publication. Both exploit weaknesses that already existed within research evaluation systems. As these threats evolve, safeguarding scientific integrity will require reforms that extend beyond detecting misconduct after publication. It will require redesigning the institutional incentives that determine what kinds of science are rewarded in the first place.
Indonesia’s experience offers a lesson for any country attempting to expand its research capacity rapidly. Bibliometric growth demonstrates scientific activity. It does not guarantee that the evidence produced is reliable. If the scientific literature increasingly contains references to studies that never existed, science risks losing its defining characteristic: the capacity of every claim to be traced back to verifiable evidence. That is not a publishing problem. It is a problem for the credibility of science itself.
Based on Brief Communication SSRN Preprint:
Ridlo, Ilham and Nurhasim, Ahmad, Paper Mills, AI-Assisted Citation Fabrication, and the Integrity of Indonesia’s Research Evidence Base (June 16, 2026). Available at SSRN: https://ssrn.com/abstract=6950579
