When a Paper No Longer Means Science: Biology’s reckoning With AI-Generated Research

A quiet but profound rupture has opened at the heart of modern science, and it is not about what artificial intelligence can do, but about what a scientific paper is supposed to mean. In a perspective article published in PLOS Biology, Ran Blekhman of the University of Chicago argues that generative AI has severed the long-standing connection between producing a manuscript and actually doing science. For more than a century, the paper served as the artifact through which biologists were assessed, funded, hired, and promoted. That assumption, he contends, no longer holds, and the field must now answer a question it has long avoided: what is biology research actually for?

The scale of the disruption is visible in the numbers. Monthly submissions to arXiv, the preprint server that has long tracked the pulse of scientific output, have nearly doubled since late 2022, an inflection point that coincides precisely with the public rollout of the first widely used generative AI systems. While some of the surge reflects growth trends that predate AI, the acceleration of the past four years is difficult to explain without invoking automated text generation. arXiv has responded with increasingly strict measures: in 2025 it stopped accepting review and position papers in computer science that had not already passed peer review, and in 2026 it announced a one-year submission ban for authors whose papers contain unchecked AI content, such as hallucinated references or leftover chatbot instructions. Blekhman regards these as reasonable but ultimately superficial responses, addressing symptoms rather than the underlying problem.

The core of the problem, as he puts it bluntly, is that it is now straightforward to generate a plausible scientific paper without doing any actual research. This is not merely an inconvenience for overworked editors and reviewers. It represents a structural collapse in the signal that the scientific record was designed to carry. A paper was never a perfect proxy for scientific value, but it was at least expensive to produce, and that cost served as evidence that some genuine work had been done. Generative AI has removed that constraint entirely, and with it, the last vestige of assurance that a manuscript corresponds to real scientific activity.

To find a way forward, Blekhman draws on an argument from the mathematician Terence Tao, who wrote in a recent essay about mathematics in the age of AI that a field grappling with AI-generated results must first state plainly what the goals of its research actually are. Blekhman applies the same exercise to biology and offers five purposes as a starting point for the conversation. First, biology research exists to explain how living systems work at every scale, from molecules to ecosystems. Second, it exists to improve human health, whether directly through therapy or through the longer chain that runs from a basic finding to a medicine decades later. Third, it builds the resources that facilitate the field itself, from datasets and model organisms to reference databases, methods, and software. Fourth, it trains the next generation of scientists who will carry the work forward. And fifth, it maintains the trust of a public that pays for almost all of it, a goal Blekhman identifies as the least frequently named but arguably the most important.

For most of the twentieth century, producing a scientific paper advanced at least some of these goals simultaneously. A paper describing a mechanistic finding also generated a dataset, trained a student in its methods, and potentially laid the groundwork for eventual therapeutic applications. Because manuscripts reflected real progress toward the field’s goals, assessment systems gradually began relying on them as primary metrics. Review panels, search committees, and promotion decisions collapsed scientific quality into manuscript counts, venue metrics, and citation tallies. Blekhman is careful to note that he is not defending this proxy, which he acknowledges is obviously flawed and had been challenged repeatedly by the late 2000s. Still, this style of evaluating science took hold because a paper required effort to produce, and that effort served as evidence that at least one goal on the list had been advanced.

Generative AI has broken this already-flawed proxy completely. A plausible paper can now be produced in a day, and the process of producing it does not necessarily enhance knowledge, build a resource, train a student, or advance any other goal of biology research. Blekhman characterizes this as a crisis that the field has not yet begun to grapple with. Some might object that scientific manuscript quality has always been uneven, that paper mills, fabricated data, and low-quality publications have been consistent features of the record long before AI arrived. But Blekhman draws a clear distinction: producing a paper, even a marginal one, still required some labor and resources, indicating some level of genuine scientific engagement. Even a weak paper likely contributed to some of biology’s goals. Automated text generation eliminates this link between manuscript production and authentic academic endeavor entirely.

Another objection holds that the surge in AI-generated manuscripts in biology is smaller than in fields such as machine learning or mathematics, so perhaps the crisis is overstated. Blekhman concedes this is partly true, because AI has not yet substantially increased productivity in biology, where the bottlenecks are structural and experimental rather than computational. The surge is nevertheless just beginning. Computational biology already sees it, because AI can fabricate a computational paper end to end. And Blekhman predicts the situation will worsen, because in biology research there are no brakes on the system. Scientists will keep generating more papers because their careers depend on it. Publishers will keep publishing them and collecting fees to increase their revenue. And AI companies will happily keep selling the tokens that enable the entire cycle.

The response, Blekhman argues, cannot be to stop evaluating scientists altogether. In the absence of a deliberate response, his prediction is that evaluation will fall back on older, cruder proxies such as journal prestige and impact factor, which would undo the past decade of research assessment reform. The San Francisco Declaration on Research Assessment and the Leiden Manifesto pushed the field to move away from journal name and citation counts as measures of quality, and made real, if slow, progress. AI-driven uncertainty about what a paper represents threatens to reverse that progress entirely. Instead, Blekhman suggests the field could use this surge as an opportunity to acknowledge that paper counts are meaningless and to abandon the current incentive system in favor of returning focus to the substance of research itself.

Goodhart’s law, often stated as the observation that when a measure becomes a target it stops being a good measure, looms over the entire discussion. Publishing more papers has been a target in biology for decades, and the consequences of that misalignment are now amplified by tools that can satisfy the target without any of the underlying work. Blekhman insists that defining new ways to evaluate science aligned with the field’s actual goals is work for the entire community: journals, funders, universities, and institutions all have a role to play. The alternative is a slow drift back toward metrics that reward volume over substance, at precisely the moment when volume has become the easiest thing to manufacture.

Perhaps the most striking aspect of Blekhman’s argument is its framing not as a call for new detection tools or stricter policies, but as a fundamental question about purpose. The field, he suggests, should begin by asking itself what biology research is for, and then rebuild its evaluation systems around the answer. Whether the community can muster the collective will to undertake that kind of self-examination, at a moment when the pressures of career advancement and institutional revenue push in the opposite direction, remains an open question. What is clear is that the old bargain, in which the effort of writing a paper served as a rough guarantee of scientific substance, has collapsed, and something must take its place.

Subject of Research: The impact of generative AI on the purpose and evaluation of biology research

Article Title: What is biology research for?

Article References: Blekhman, R. (2026). What is biology research for?. PLOS Biology, 24(10), e3004075. https://doi.org/10.1371/journal.pbio.3004075

Image Credits: AI Generated

DOI: 10.1371/journal.pbio.3004075

Keywords: generative AI, scientific publishing, research assessment, PLOS Biology, Ran Blekhman, arXiv, Goodhart's law, research integrity, biology, Terence Tao, impact factor, peer review