You can feel it in the day-to-day rhythm of academic life: the inbox full of “special issue” invitations, the relentless grant cycles, the performance reviews that quietly translate curiosity into counts. Many of us have a nagging suspicion that science is becoming more productive and less interesting — that we are publishing more while discovering less.
On paper, the system looks healthy. Worldwide publication output is enormous and still rising. The U.S. National Science Foundation’s Science & Engineering Indicators puts global S&E article output at about 3.3 million articles in 2023 (Scopus-indexed), up from roughly 2.2 million in 2014 [1]. That is not a gentle increase; it is a regime change. And once the system scales that fast, the uncomfortable question is no longer “Are we doing science?” It is: “What, exactly, are we optimising?”
Because what we often call “excellence” has quietly been redefined. Publications, citations and grant income were meant to be imperfect signals of contribution. Instead, they have become the job description. When metrics become targets, researchers behave like rational adults in an irrational system: split studies into smaller outputs, prioritise what will publish quickly, and avoid projects that might fail slowly. The literature becomes a production line of risk-managed novelty. Commercial publishing models that profit from throughput do not need to corrupt science; they merely need to align with the incentives already in place [8]. The Leiden Manifesto warned about this years ago: metrics are useful servants and disastrous masters [9].
Once you reward volume, you should not be surprised when the ecosystem produces volume — including the kind you don’t want. That is where the conversation turns from awkward to ugly: manipulated peer review, ghost authorship, paper mills. These are not new infections introduced by AI; they are opportunistic pathogens thriving in a weakened immune system [4]. What has changed is the scale. Major publishers have had to retract thousands of papers and even shut or restructure journals under paper-mill pressure; Wiley’s Hindawi portfolio is the notorious example, with over 11,000 retractions reported and multiple journals discontinued as part of integrity remediation [5]. When a publisher reaches five-figure retractions, that is not a “few bad apples”. That is an industry problem.
Retractions are a blunt instrument, and yes, better detection contributes. But even if you take the most charitable interpretation, the trend still screams: the quality-control layer is not keeping up. CNRS recently summarised the scale starkly, pointing to around 10,000 retracted articles in 2023 against a backdrop of roughly 3 million new publications per year [2]. In other words, we are flooding the marketplace of ideas so fast that even low failure rates become a high-volume disaster — enough to waste careers, distort fields, and quietly train early-career researchers to treat the literature as something you skim, not something you trust.
And beneath integrity sits the deeper fear: even when papers are “valid”, are they doing anything intellectually brave? One of the most provocative empirical findings in recent scientometrics is Park, Leahey and Funk’s 2023 Nature analysis, which examined tens of millions of papers and patents and concluded that research outputs have become less disruptive over time [3]. That does not mean scientists are dumber. It means the system is increasingly structured to reward consolidation, extension, and incremental additions — the safe middle lane of knowledge production. We are building taller towers on the same foundations, and calling it progress because it is measurable.
This is where the “Nobel argument” becomes emotionally persuasive. Earlier eras are remembered for conceptual revolutions — quantum mechanics, DNA, plate tectonics — while contemporary breakthroughs often arrive through long arcs of cumulative work, massive infrastructure, and large teams. That shift is not necessarily bad; complexity is real. But it does expose a brutal truth: if you want a career, you are often better off being reliable than being right in a genuinely new way. The system says it loves breakthrough science; it mostly funds deliverables.
The university has reorganised itself accordingly. Senior academics increasingly operate as portfolio managers: budgets, staffing, compliance, stakeholder engagement, industry partnerships, reporting. None of this is inherently evil, but it changes what gets rewarded. When promotion depends on KPIs and income streams, professors become managers who outsource intellectual risk downward. The lab becomes a small enterprise. Students become workforce. And “mentorship” can quietly mean “keep the machine running”.
Industry collaboration sits uncomfortably inside this ecosystem. At its best, it is essential: it scales ideas, builds capabilities, and translates discoveries into public benefit. At its worst, it acts as a philosophical solvent. Companies are designed to produce products and profit. Universities are increasingly pressured to behave the same way. The predictable result is a research culture that treats PhDs as a supply chain and calls it “impact”. The risk is not that applied research exists; it is that fundamental science becomes the hobby you pursue after the quarterly report.
So where do LLMs fit? Not as the original sin, but as a stress test — and, frankly, an embarrassment. The publication explosion, peer-review overload, paper mills and integrity crises were already underway before generative AI became mainstream [2,4]. LLMs simply make it cheaper to generate plausible academic text at scale. If your evaluation system cannot distinguish deep insight from fluent filler, that is not an AI problem. That is a governance problem. AI did not ruin science; it revealed how much of “scientific communication” had already drifted towards performative output.
There are, of course, counterarguments that deserve respect. Science is more global, more collaborative, and more technically powerful than ever. The same period that produced metric pressure also produced extraordinary capabilities: high-throughput sequencing, climate modelling, synchrotron science, advanced microscopy, computational methods that earlier generations could not imagine. Some apparent “decline” may simply reflect maturity: when fields are large, progress becomes specialised and cumulative. Even the disruption study cautions against nostalgia; modern frontiers often advance through collective builds rather than solitary leaps [3].
Still, the incentive problem does not disappear just because science is powerful. We have built an ecosystem where the safest way to survive is to publish frequently, cite strategically, and avoid ideas that might fail slowly. If that diagnosis is even partly correct, the solution is not to ban AI or shame researchers. It is to change what the system pays for.
If I had to compress the reform agenda into one sentence, it would be this: stop treating bibliometrics as goals and start treating them as noisy signals. That is the spirit of the Leiden Manifesto: evaluate with expert judgement, avoid false precision, and align indicators with the mission of science rather than the convenience of accounting [9]. Reward fewer, stronger contributions. Make replication and negative results prestigious. Credit peer review. Harden editorial checks. Protect long-horizon fundamental work from the tyranny of short-cycle deliverables.
And perhaps the final provocation: we may be watching an evolutionary transition. Machines absorbed physical labour, and society reorganised around it. Now machines may absorb routine cognitive production — drafting, summarising, coding, even parts of analysis. The question is whether humans use that shift to reclaim depth, or whether we simply produce more paperwork faster and call it “innovation”. The future of science is unlikely to be decided by AI. It will be decided by what we choose to reward when AI makes “more” effortless.
Written by Vitaliy Ponomar
Bibliography
[1] National Science Board (2024). Science and Engineering Indicators 2024. National Center for Science and Engineering Statistics (NCSES), U.S. National Science Foundation. Arlington, VA.
[2] Centre National de la Recherche Scientifique (CNRS) (2024). Scientific publications: fatal level of overproduction? CNRS News.
[3] Park, M., Leahey, E., & Funk, R. J. (2023). Papers and patents are becoming less disruptive over time. Nature, 613, 138–144.
[4] Retraction Watch Database (Center for Scientific Integrity). Ongoing dataset of retracted scientific publications.
[5] Wiley (2023–2024). Public statements regarding Hindawi journal retractions and portfolio restructuring due to paper-mill activity.
[6] Fanelli, D. (2018). Opinion: Is science really facing a reproducibility crisis, and do we need it to? Proceedings of the National Academy of Sciences, 115(11), 2628–2631.
[7] Ioannidis, J. P. A. (2005). Why most published research findings are false. PLoS Medicine, 2(8), e124.
[8] Larivière, V., Haustein, S., & Mongeon, P. (2015). The oligopoly of academic publishers in the digital era. PLoS ONE, 10(6), e0127502.
[9] Hicks, D., Wouters, P., Waltman, L., de Rijcke, S., & Rafols, I. (2015). The Leiden Manifesto for research metrics. Nature, 520, 429–431.