The AI is not lying intentionally. It is simply making things up.
In the vast architecture of human knowledge, trust has always been the mortar between the bricks — and now that mortar is quietly being replaced with something that only resembles it. Artificial intelligence systems, increasingly woven into academic writing, are generating citations that point to nowhere: invented studies, phantom authors, journals that never carried the work in question. Measured across biomedical literature, the contamination already touches thousands of papers, and the fields moving fastest toward AI adoption are showing the deepest cracks. The question this moment poses to scholarship is ancient even if the mechanism is new: how do we preserve the integrity of a shared record when the tools we use to build it cannot distinguish the real from the plausible?
- AI writing tools are fabricating citations at a measurable and growing scale — roughly one in every 277 PubMed-indexed papers from 2026 contains at least one reference that does not exist.
- Social sciences preprints, posted before peer review, show the highest concentration of hallucinated references, exposing a fault line between fields with robust quality controls and those without.
- The danger compounds silently: a researcher trusts AI output, publishes a false citation, a reader cites it in turn, and a fiction travels deeper into the permanent literature with each iteration.
- Peer review, already strained, cannot realistically verify every reference at scale, and journals scrambling to write AI-use policies are racing against a technology that is evolving faster than the rules meant to govern it.
- Detection tools that cross-check citations against real databases are emerging, but the field faces a structural tension that no single safeguard resolves: AI is genuinely useful for writing, yet constitutionally indifferent to factual truth.
Something strange is spreading through the academic literature. Artificial intelligence systems, when used to draft or assist with research papers, are inventing citations — conjuring studies that never existed, attributed to researchers who may never have written them, placed in journals that may never have published them. The scale is now measurable: roughly one in every 277 papers indexed in PubMed in 2026 contains at least one fabricated reference. Translated into absolute numbers, that means thousands of papers, each one capable of sending readers down a trail of evidence that leads nowhere.
The behavior stems from how large language models are built. These systems learn to predict plausible sequences of text — and a citation is, to them, simply a pattern to reproduce. They generate author names, journal titles, and publication years that look correct because they have learned what citations look like, not because they have verified that any particular citation exists. Researchers call this hallucination: not deliberate deception, but confabulation at scale.
The problem is unevenly distributed. Social sciences preprints — papers posted online before peer review — show the highest rates of fabricated references, likely because AI writing tools are being adopted quickly there and pre-publication scrutiny is limited. Biomedical research fares somewhat better, but even a single hallucinated citation in a medical paper can misdirect clinicians or researchers toward evidence that does not exist.
Underlying all of this is a broken assumption: that citations in published papers have been checked. As AI tools enter the writing process, that assumption no longer holds. A researcher may accept AI-drafted literature reviews without verification; a reader may then cite those same phantom references, carrying the fiction further into the record.
Journals are beginning to respond — requiring authors to certify citation accuracy, restricting AI use in certain manuscript sections — but enforcement is difficult and the technology moves faster than policy. Some institutions are building detection tools that scan papers against real databases to flag nonexistent references. The deeper tension, however, remains unresolved: AI is genuinely useful for organizing and drafting text, yet it is structurally unreliable when factual accuracy is what the work demands. How the research community navigates that contradiction will determine whether the foundation of shared knowledge holds.
Something strange is happening in the academic literature. Researchers are discovering that artificial intelligence systems, when asked to write papers or help draft them, are inventing citations wholesale—conjuring up studies that never existed, authored by researchers who may never have written them, published in journals that may not have carried them. The problem has grown large enough to measure. An analysis of papers indexed in PubMed, the massive database of biomedical literature, found that roughly one in every 277 papers published in 2026 contains at least one fabricated reference. That may sound like a small fraction until you do the math: thousands of papers, each one potentially misleading readers down a false trail of evidence.
The issue traces back to how large language models work. These AI systems are trained on vast amounts of text and learn to predict what words should come next in a sequence. When asked to cite a source, they don't actually retrieve anything from a database. Instead, they generate text that looks like a citation—a plausible author name, a plausible journal title, a plausible year. The system has learned what citations look like, but it has no mechanism to verify that what it produces actually exists. Researchers call this "hallucination," a term that captures something both mundane and unsettling: the AI is not lying intentionally. It is simply making things up because that is what it does.
The problem is not evenly distributed across fields. Social sciences preprints—papers posted online before peer review—show the highest rates of fabricated citations. This may reflect the fact that researchers in those fields are adopting AI writing tools earlier, or that the papers are less likely to be scrutinized before posting, or both. Biomedical research, which has more established quality-control mechanisms, shows lower rates, though the absolute numbers remain troubling. A single hallucinated citation in a medical paper can send clinicians or other researchers down a dead end, wasting time and potentially influencing treatment decisions based on evidence that does not exist.
The discovery has surfaced a deeper problem: most researchers do not systematically verify their citations. They trust that if a citation appears in a paper, someone has checked it. But as AI tools become more common in the writing process, that assumption is breaking down. A researcher might use an AI system to draft a literature review, accept most of the output without checking, and unknowingly publish fabricated references. The reader, in turn, trusts the citation and may cite it themselves, spreading the false reference further through the literature.
The implications are significant. Academic knowledge is supposed to build on a foundation of verified facts and documented evidence. When citations become unreliable, that foundation cracks. Peer reviewers cannot catch every hallucinated reference—they would need to verify every single one, a task that is already time-consuming and is becoming impossible at scale. Journals are beginning to implement new policies: some now require authors to certify that they have verified all citations, or they prohibit the use of AI in certain parts of the manuscript. But enforcement is difficult, and the technology is advancing faster than the safeguards.
What happens next depends partly on how seriously the research community takes the problem. Some institutions are developing tools to detect hallucinated citations, scanning papers against actual databases to flag references that do not exist. Others are pushing for stronger norms around AI use in academic writing. But the underlying tension remains: AI systems are genuinely useful for drafting and organizing text, yet they are fundamentally unreliable when it comes to factual accuracy. The question is not whether to use them, but how to use them without letting false information seep into the permanent record of human knowledge.
Citas Notables
Researchers call this 'hallucination'—the AI is not lying intentionally, but simply making things up because that is what it does.— Academic researchers studying AI citation fabrication
La Conversación del Hearth Otra perspectiva de la historia
Why does this happen? Can't the AI just look up the citation?
That's the thing—it doesn't look anything up. It predicts what text should come next. It has learned what citations look like from training data, so it generates plausible-sounding ones. It's not consulting a database. It's guessing.
But researchers must notice when they cite something that doesn't exist, right?
You'd think so. But most people don't verify every citation in their own paper. They trust that if they wrote it down, it's real. And when an AI drafts the citations for them, they're even less likely to check.
So this is a problem of trust breaking down?
Exactly. The whole system depends on citations being reliable anchors. Once they become unreliable, readers can't trust the foundation of the argument.
Why are social sciences preprints hit hardest?
Probably because those fields are adopting AI tools faster, and preprints skip peer review. There's less friction before the paper goes public.
What's the fix?
That's the hard part. You could require verification, but that's labor-intensive. You could build detection tools, but they're always playing catch-up. Or you could change how we use AI—keep it out of the citation process entirely.
Will that actually happen?
Slowly, probably. Institutions are starting to set policies. But the technology is moving faster than the guardrails.