The engagement with literature is becoming increasingly superficial
In the age of generative AI, the ancient practice of citation — the careful threading of one mind's work through another's — is quietly unraveling. A Lancet study tracing more than 97 million references has found that fabricated citations in academic papers increased sixfold between 2023 and 2025, with the pace still accelerating into 2026. What began as an embarrassing near-miss for one researcher has illuminated a systemic fracture: the tools built to accelerate knowledge are, in some hands, hollowing out its foundations.
- Hallucinated citations — references to papers that never existed — have surged from 1 in 2,828 papers in 2023 to 1 in 277 in early 2026, a trajectory that alarms researchers who depend on citations to trace the lineage of scientific ideas.
- More than a third of all fabricated references cluster around just two large open-access publishers, suggesting that volume-driven, fee-for-publication models may be creating conditions where speed overtakes scrutiny.
- Experts warn the damage runs deeper than bad footnotes: systematic reviews and clinical guidelines built on phantom citations can send researchers — and eventually patients — chasing ghosts through a corrupted archive.
- Major journals including Science, NEJM, and JAMA have deployed automated verification systems, but the tools are imperfect, risking false positives that flag legitimate references with formatting quirks or non-English sources.
- Beneath the technical problem lies a cultural one: researchers under relentless publication pressure are outsourcing literature engagement to AI, trading depth for speed in ways that may quietly degrade the quality of science itself.
A Columbia University nurse and health AI researcher named Maxim Topaz nearly published a citation to a paper that had never existed. He had used an AI chatbot to help refine an editorial, checked the references himself, and still missed it — the journal's editor caught it first. That moment of mortification became the seed of a larger investigation.
Topaz and his team analyzed more than 2 million papers and 97 million citations, finding roughly 4,000 fabricated references across some 2,800 papers. The raw number might seem small, but the trajectory is stark: fabricated citations appeared in 1 of every 2,828 papers in 2023, 1 in 458 by 2025, and 1 in 277 in the first seven weeks of 2026 alone. The Lancet study, published Thursday, points squarely at generative AI as the likely engine of this acceleration.
Citations are meant to function as a genealogy of ideas — a traceable lineage anchoring new work in what came before. A hallucinated reference is not a typo; it is a phantom in the scientific archive. When researchers build systematic reviews or clinical guidelines, they rely on that record. Fabricated citations can send them chasing studies that were never written.
The problem is concentrated in certain corners of publishing. More than a third of all fabricated citations originate from just two large open-access publishers whose revenue depends on processing high volumes of submissions. Topaz declined to name them, but the pattern raises uncomfortable questions about what incentives volume-driven publishing creates.
Northwestern's Mohammad Hosseini frames the citations as a symptom of something deeper: researchers once read papers deliberately before citing them. Now many simply prompt an AI and fold the results into a draft. 'The engagement with the literature is becoming increasingly more superficial,' he said, 'and that is neither good for the researcher, nor for society.' Publication pressure accelerates the dynamic — AI promises speed, and speed is what the system rewards.
Some major journals have responded with automated verification tools, and the Science family reports catching no fabricated citations in published work. But the fix carries its own risks: automated checkers can flag legitimate references that contain formatting errors or come from databases the tools index poorly. The broader question, as sociologist Misha Teplitskiy of the University of Michigan put it, is whether AI is making science faster without making it better — and whether the publishing system can adapt before the damage compounds.
A nurse and health AI researcher at Columbia University named Maxim Topaz had an embarrassing near-miss. He used an AI chatbot to help refine an editorial for journal submission, checked the citations carefully, and still almost published a reference to a paper that did not exist. The journal's editor caught it. That moment of mortification became the seed for a larger investigation—one that has now revealed something troubling about the state of scientific publishing in the age of generative AI.
Topaz and his team sifted through more than 2 million academic papers and 97 million citations using automated detection tools. They found approximately 4,000 fabricated citations scattered across roughly 2,800 papers. The raw count is small enough that some might dismiss it. But the trajectory is what matters. In 2023, fabricated citations appeared in roughly 1 out of every 2,828 papers. By 2025, that ratio had shifted to 1 in 458 papers—a sixfold increase in just two years. In the first seven weeks of 2026 alone, the rate had accelerated further to 1 in 277 papers. The Lancet study, published Thursday, documents this acceleration and points squarely at generative AI tools as the likely culprit.
Citations are supposed to function as a kind of genealogy for ideas. They anchor a researcher's work in the literature that came before, creating a traceable lineage of knowledge. A fabricated citation—a reference to a paper that was never written, never published, never existed—corrupts that record. It is not merely a typo or a formatting error. It is a phantom in the scientific archive. When researchers build systematic reviews or develop clinical guidelines, they rely on citations to identify relevant studies. A hallucinated reference can send them chasing ghosts. More broadly, it pollutes the public record of science itself.
The problem appears concentrated in certain corners of the publishing landscape. More than a third of all fabricated citations come from just two publishers, both large open-access operations that charge authors substantial fees to publish without paywalls. Topaz declined to name them publicly, but the pattern is clear: revenue-driven publishing models may create perverse incentives. When a publisher's income depends on volume—on the number of papers it can move through the system—the pressure to process submissions quickly can outpace the rigor needed to catch AI-generated hallucinations.
Mohammad Hosseini, a professor at Northwestern University who studies research integrity, sees the fabricated citations as a symptom of a deeper cultural shift. Researchers once engaged with literature deliberately. They would read a paper, take notes, and then decide whether it belonged in their own work. Now, many simply feed a prompt to ChatGPT or another large language model and sprinkle the resulting citations into their draft. It is faster. It requires less thought. And it works—until it does not. "The engagement with the literature is becoming increasingly more superficial," Hosseini said, "and that is neither good for the researcher, nor for society, nor for our publication practices." The shift reflects something else too: the desperation to publish. Researchers face immense pressure to accumulate publications, and AI tools promise to accelerate that process. The result is a kind of arms race toward sloppiness.
Some major journals have begun to fight back. The Science family of journals uses automated tools to validate references and reports having caught no fabricated citations in published papers. The New England Journal of Medicine and JAMA have similar systems in place. The Public Library of Science acknowledged seeing "numerous" unverifiable references in submissions and said it is exploring ways to incorporate citation verification into its workflows. But the challenge is real: automated tools can flag legitimate references as false if they contain formatting errors, incomplete information, or if the source material is published in a language the database does not index well. The cure, in other words, risks creating its own problems.
Misha Teplitskiy, a sociologist of science at the University of Michigan, called the Lancet study "one of the first papers that's telling us something about the quality of what's being produced with LLMs." For months, researchers have published studies tracking how academics use large language models in their work. Few have asked whether that work is actually any good. This paper does. And the answer, so far, is that AI is making science faster but not necessarily better—and in some cases, it is making it worse. The question now is whether the publishing system can adapt quickly enough to catch up with the tools that are reshaping how research gets done.
Citações Notáveis
This is one of the first papers that's telling us something about the quality of what's being produced with LLMs, and it's a signal of slop.— Misha Teplitskiy, sociologist of science at University of Michigan
The engagement with the literature is becoming increasingly more superficial, and that is neither good for the researcher, nor for society, nor for our publication practices.— Mohammad Hosseini, professor at Northwestern University studying research integrity
A Conversa do Hearth Outra perspectiva sobre a história
Why does a fabricated citation matter so much? It's just a wrong reference.
Because citations are how science builds on itself. If I cite a study that doesn't exist, someone else might try to build on it too. It corrupts the entire chain.
But you said the actual number of fabricated citations is low—4,000 out of 97 million.
The number is low, yes. But it's growing sixfold in two years. That's the alarm. It's not the current state; it's the trajectory.
So this is really about AI making things faster without making them better?
Faster and lazier. A researcher used to read a paper, think about it, decide if it mattered. Now they prompt ChatGPT and take whatever comes back. The AI sometimes invents citations that sound plausible.
Why would publishers let this happen?
Some of them profit from volume. Open-access publishers charge authors to publish. More papers published means more revenue. Speed becomes more important than accuracy.
Can't journals just check the citations?
They're trying. But automated tools have false positives—they flag real papers as fake because of formatting issues or language barriers. It's harder than it sounds.
What happens if this keeps accelerating?
Systematic reviews and clinical guidelines built on these papers become unreliable. And the culture of reading and thinking about research gets replaced by box-checking.