Imagine science as a grand, interconnected family tree, where each new discovery builds upon the wisdom of those that came before. Citations in academic papers are like the meticulous records in this family tree, showing us the lineage of ideas, the roots of experiments, and the people who laid the groundwork for our understanding. They’re meant to connect us to the past, grounding new research in a rich tapestry of established knowledge. But lately, something unsettling has been happening. A creeping problem, like a gnawing pest, is starting to unravel this carefully constructed fabric. More and more often, these academic breadcrumbs lead to nowhere – to papers that simply don’t exist. These “fabricated” citations, as a new study in The Lancet tragically reveals, are seeping into the very core of scientific literature, muddying the waters of public knowledge. And while the full extent of the damage is still unfolding, the finger of blame is increasingly pointing at a powerful new tool: generative artificial intelligence. These AI programs, while promising efficiency, are inadvertently creating a “slop” that contaminates the wellspring of scientific truth.
This isn’t just a distant academic concern; it’s a growing frustration for researchers worldwide. Picture this: you receive an exciting alert, indicating that your work has been cited in a new publication. A thrill runs through you – validation, recognition! But then, as you go to investigate, a cold splash of reality hits. The paper you’re supposedly cited in? It’s a phantom, an illusion. This experience is becoming alarmingly common. Last year, even a high-profile report from the White House, detailing its strategic priorities for tackling chronic diseases, was found to contain several incorrect citations. The suspicion? Artificial intelligence, once again, was the likely culprit. This widespread phenomenon caught the attention of Maxim Topaz, a nurse and health AI researcher at Columbia, who spearheads this very investigation. His own wake-up call to this looming crisis was deeply personal and embarrassing. He had, quite innocently, used an AI chatbot to help refine an editorial he was submitting to a journal. Despite his diligent checks, an astute editor at the journal still managed to flag an erroneous citation. “I was deeply embarrassed,” Topaz confessed, “I checked for that, and it still almost happened to me. This is how I ended up thinking about other people.” This personal brush with fabricated citations ignited his determination to quantify the problem.
And so, Topaz embarked on a monumental task, sifting through a staggering two million papers and 97 million citations, painstakingly aided by AI tools to detect the subtle whispers of deception. His findings, while initially appearing small in number – around 4,000 fabricated citations spread across 2,800 papers – reveal a far more insidious truth. The real alarm bell isn’t the current volume, but the frightening acceleration of this trend. In 2023, approximately one in every 2,828 papers contained at least one fabricated reference. A year later, in 2024, that number had plummeted to one in 458 – a staggering six-fold increase. And during the first seven weeks of 2026, the rate continued its unsettling climb, reaching an alarming one in 277 papers. This exponential growth isn’t just a statistical blip; it signifies a rapidly escalating problem that threatens the very integrity of scientific discourse. The proliferation of these phantom citations isn’t just about misinformation; it could directly compromise crucial systematic reviews and clinical guidelines, which depend on an accurate, verifiable foundation of research. More profoundly, it hints at deeper systemic issues within the scientific publication ecosystem.
Mohammad Hosseini, a professor at Northwestern University who delves into research integrity and the ethical implications of AI, articulates a profound concern stemming from these “hallucinated” citations. He observes, “It shows that there’s people who don’t even want to spend half an hour to check the references of a paper.” This isn’t merely negligence; it speaks to a deeper malaise: an overwhelming pressure to publish quickly and frequently, often at the expense of meticulous scholarship. It’s a stark reflection of a “flawed scholarly evaluation model” that excessively prioritizes peer-reviewed publications as the gold standard for career advancement. Furthermore, Hosseini believes it signals a troubling shift in the very culture of citations. What was once a thoughtful, genuine engagement with another researcher’s work, a true acknowledgment of intellectual debt, is now, for some, reduced to a mere “box-checking exercise.” The deep, reflective engagement with existing literature, where researchers would carefully consider the relevance and validity of a prior work, is steadily eroding. Now, with the ease of AI, some are simply using “hunches to prompt ChatGPT or other AI tools,” generating a list of citations to “sprinkle over their papers.” This superficial approach is detrimental to everyone involved – the individual researcher, the scientific community, and indeed, society at large, as the foundation of knowledge becomes increasingly shaky.
Interestingly, these fabricated citations aren’t uniformly distributed throughout the vast landscape of academic publishing. The researchers’ dashboard, accompanying their paper, paints a concerning picture: over a third of these manufactured references originate from just two publishers. While Topaz maintains a diplomatic silence on the specific names, he does reveal that they are large, open-access publishers. This distinction is crucial; open-access models often rely on authors paying hefty fees to publish their work, circumventing traditional paywalls. This business model, while promoting accessibility, can inadvertently create incentives for rapid publication that might compromise thorough vetting. In response to these unsettling findings, STAT, a prominent science news outlet, reached out to several leading biomedical research publishers to gauge their awareness and strategies. The ‘Science’ family of journals, for instance, employs an automated tool to scrutinize references and proudly reported not having encountered any fabricated citations in their published papers. Similarly, spokespeople for the highly esteemed ‘New England Journal of Medicine’ and ‘JAMA’ confirmed the use of sophisticated citation validation tools, emphasizing that authors are held accountable for the accuracy of their submissions. However, the ‘Public Library of Science’ (PLOS), a major open-access publisher, offered a more candid assessment, admitting to seeing “numerous” unverifiable references in manuscripts awaiting publication.
Renee Hoch, the head of publication ethics at PLOS, offers a glimpse into the complexities of addressing this issue. She acknowledges that PLOS is actively exploring ways to integrate robust checks into their publishing workflows. However, the implementation isn’t as straightforward as one might assume. During their pilot programs, they’ve encountered a significant number of “false positives,” where legitimate references were mistakenly flagged. This can happen for various reasons: incorrect or incomplete information in reference lists, formatting inconsistencies that confuse automated systems, challenges with publications in different languages, or limitations within commonly used research literature databases. This highlights the delicate balance publishers face: the urgent need to combat fabricated citations without inadvertently hindering the publication of legitimate, well-researched work. The challenge, therefore, isn’t just about identifying the fakes; it’s about developing sophisticated, nuanced tools and ethical guidelines that can safeguard the integrity of scientific knowledge in an increasingly AI-driven publishing landscape, ensuring that the grand family tree of science remains rooted in truth.