20 Fake Citations Slip Past Peer Review: AI 'Hallucinations' Expose Crisis in Academic Publishing

They looked legitimate. The titles sounded academic. The authors' names seemed plausible. There was just one problem: Twenty of the references in a peer-reviewed academic paper didn't exist at all.

The phantom citations—created by artificial intelligence "hallucinations"—sailed through the rigorous peer review process at an established academic journal, exposing a dangerous vulnerability in the system that underpins all scientific knowledge.

What a team of expert reviewers missed, a curious social media user caught. And now, the University of Hong Kong is investigating, a professor has apologized, and academic publishers worldwide are facing an uncomfortable truth: AI is breaking their quality control systems in ways they never anticipated.

The Discovery That Shouldn't Have Been Necessary

The paper seemed unremarkable at first glance. Published on October 17 in the journal China Population and Development Studies, "Forty years of fertility transition in Hong Kong" appeared to be a standard academic analysis of demographic trends—the kind of specialized research that gets published daily in thousands of journals worldwide.

But something caught the attention of a social media user who "was told by a friend" that many of the paper's references looked suspicious. A closer examination revealed the shocking truth: At least 20 out of 61 citations—nearly one-third of the paper's entire reference list—were complete fabrications.

The articles didn't exist. The journals hadn't published them. The authors were phantom names attached to phantom research. They were the product of what AI researchers call "hallucinations"—instances where artificial intelligence systems confidently generate false information that appears entirely plausible.

And they had fooled everyone involved in the publication process.

How AI Creates Convincing Lies

To understand how this happened, you need to understand how AI hallucinations work.

When researchers use AI tools like ChatGPT to help with literature reviews or generate citations, these systems don't actually search databases of real papers. Instead, they predict what a plausible citation might look like based on patterns in their training data.

The result? Citations that follow all the right formatting conventions, with realistic-sounding titles, plausible author names, and appropriate journal names. They look exactly like real academic references because the AI has learned what real references look like.

But they're entirely fictional.

The problem is particularly insidious because these fake citations are hard to spot without checking each one individually. They don't trigger obvious red flags like misspellings or formatting errors. They pass the "glance test" that busy peer reviewers and editors rely on.

"AI can generate references that look legitimate in every way except that they reference works that never existed," explains Robert Davison, a professor of information systems at City University of Hong Kong and a veteran journal editor. "Without verification, there's no way to know."

The Peer Review Failure

The incident raises uncomfortable questions about academic publishing's most sacred safeguard: peer review.

In theory, peer review is academia's quality control system. Experts in the field examine submitted papers before publication, checking methodology, verifying claims, and—in theory—confirming that citations actually exist and support the arguments being made.

In practice, peer reviewers are unpaid volunteers juggling their own research, teaching, and administrative duties. They're evaluating papers in their spare time, often reviewing several per month. The expectation that they'll verify dozens of citations individually is, frankly, unrealistic.

"Enforcement ultimately depends on editors, who must carefully verify reference lists," Davison notes. But even this safeguard has limits. Journal editors are even more overwhelmed than reviewers, processing hundreds of submissions annually.

The University of Hong Kong paper exposed the gap between theory and reality. Professional academics with expertise in demographics and population studies reviewed the paper. They approved it for publication. And they never caught that one-third of the references were fake.

It took a tip from a friend to a social media user to reveal what the experts missed.

The Apology and the Defense

Professor Paul Yip Siu-fai, the paper's corresponding author and a faculty member in HKU's social work and social administration department, issued an apology on Sunday on behalf of himself and his PhD student, Bai Yiming.

The corresponding author role carries special responsibility—it signifies the person who stands behind the work's integrity and serves as the primary contact for questions about the research. By putting his name on the paper, Yip was vouching for its accuracy.

The apology acknowledged the failure but didn't detail how the fake citations ended up in the paper or what AI tools might have been used.

Meanwhile, the journal itself mounted a defense that reveals much about how academic publishing is struggling to adapt to the AI era.

China Population and Development Studies insisted that the paper's "core conclusions remained valid despite 'some mismatches and inaccuracies' with the citations."

Read that again: "Some mismatches and inaccuracies."

One-third of the references being completely fabricated isn't a mismatch or an inaccuracy. It's a fundamental failure of scholarly integrity. Yet the journal's response suggests a reluctance to grapple with the severity of the problem.

This defensive crouch isn't surprising. Acknowledging the full scale of the AI citation crisis would require journals to confront uncomfortable truths about their own quality control systems—and potentially retract countless papers while investigating whether other published works contain similar problems.

The Systemic Crisis No One Saw Coming

The Hong Kong incident isn't isolated. It's emblematic of a larger crisis brewing in academic publishing.

"While major journals typically have their own policies on the use of artificial intelligence for academic work, enforcement ultimately depends on editors," Davison warns. The policies exist on paper, but the practical mechanisms for enforcement are overwhelmed by the scale and sophistication of AI-generated content.

Consider the math: Thousands of academic journals publish millions of papers annually. Each paper contains dozens of citations. If even a small percentage of researchers are using AI tools carelessly—or deliberately exploiting them—the number of fake citations already in the published record could be staggering.

And we have no systematic way to find them.

Unlike plagiarism, which can be detected with software that compares text against databases, fake citations don't match anything because they don't exist. They're invisible to automated detection systems. Finding them requires manual verification—checking each citation individually to confirm the referenced work actually exists.

For a journal publishing hundreds of papers per year, each with 50-100 citations, that's tens of thousands of references to verify manually. It's simply not feasible with current resources.

What This Means for Trust in Science

The implications extend far beyond academic publishing.

Scientific knowledge is cumulative. Researchers build on previous work, citing earlier studies to support new hypotheses and findings. This citation network is the backbone of scientific progress—it's how we know what we know.

When fake citations enter this network, they corrupt the knowledge base. A researcher might cite a paper that cites a fake reference, creating chains of scholarship built partially on fiction. Over time, these phantom citations can become embedded in the literature, referenced repeatedly without anyone checking if the original source actually exists.

The Hong Kong paper was about fertility trends—relatively low-stakes research in terms of immediate real-world impact. But what about medical research that informs treatment decisions? Environmental studies that shape policy? Engineering papers that guide infrastructure design?

If AI-generated fake citations can slip into demographic research, they can slip into anything.

The Enforcement Gap

Davison's call for "better enforceable policies to regulate the role of artificial intelligence in academic work" highlights the central challenge: policies without enforcement mechanisms are just suggestions.

Many journals now have AI policies. They typically require authors to disclose AI use and prohibit using AI to generate core content. But these policies rely on the honor system. There's no practical way to verify compliance, and violations only come to light when someone—usually by accident or through a tip—discovers the fraud.

The Hong Kong case demonstrates the problem perfectly. The paper made it through submission, peer review, editing, and publication. It was only caught because a social media user followed up on a friend's suspicion.

This can't be the system we rely on.

What Needs to Change

Addressing the AI citation crisis will require multiple approaches:

Technological solutions: Academic publishers need to invest in automated systems that verify citations against databases of real publications. Some tools already exist, but they're not universally implemented.

Cultural shifts: The academic culture of "publish or perish" creates pressure to produce papers quickly, which incentivizes corner-cutting and careless AI use. Slowing down and prioritizing quality over quantity would help.

Editorial resources: Journals need more support for editors and reviewers. This likely means paid positions, more rigorous training on AI-related issues, and smaller reviewer workloads that allow thorough examination.

Transparency requirements: Detailed disclosure of exactly how AI was used in research—not just whether it was used—would help reviewers identify high-risk papers that need extra scrutiny.

Consequences: Clear penalties for submitting papers with fake citations, whether intentional or careless, would create stronger incentives for careful verification.

But implementing these changes requires resources that many journals—particularly smaller, specialized publications—simply don't have. And it requires acknowledging the scale of the problem, which many in academic publishing remain reluctant to do.

The Question Everyone Should Be Asking

Here's the most unsettling part of the Hong Kong incident: It was discovered by accident.

A friend noticed something odd. They mentioned it to someone else. That person checked. And they found that one-third of the citations were fake.

How many other papers have similar problems but haven't been discovered? How many fake citations are already embedded in the academic literature, being cited in new papers, building scaffolding for research built partially on fiction?

We don't know. And we have no systematic way to find out.

The University of Hong Kong is investigating this single case. Professor Yip has apologized. The journal has defended its publication. But none of this addresses the larger question: If peer reviewers, editors, and academic institutions couldn't catch this problem—and only caught it because of a tip from a friend—how can we trust any of it?

The answer is increasingly unclear. And that should worry everyone who cares about the integrity of scientific knowledge.

A Warning for the Future

The AI hallucination crisis in academic publishing is still in its early stages. As AI tools become more sophisticated and more widely used, the problem will likely get worse before it gets better.

These tools aren't going away. Researchers will continue using them because they're useful—they can speed up literature reviews, help with formatting, and assist with language in ways that benefit non-native English speakers. The problem isn't AI use itself; it's careless or malicious AI use combined with inadequate verification systems.

The Hong Kong incident is a warning shot. Twenty fake citations in one paper exposed the vulnerability. How many more papers are out there with similar problems? How many phantom references are already cited in other works, creating cascading chains of fictional scholarship?

Until academic publishing develops robust systems for detecting and preventing AI-generated fake citations, every paper is potentially suspect. Every reference list needs verification. Every claim built on citations requires double-checking.

In an era when misinformation is already eroding trust in expertise and institutions, academia can't afford to have its quality control systems compromised by AI hallucinations.

But that's exactly what's happening. And one social media user's casual investigation just proved it.

The University of Hong Kong's investigation is ongoing. The journal has not announced plans to retract the paper, maintaining that its core conclusions remain valid. The incident has sparked calls from veteran editors for stronger, enforceable AI policies in academic publishing.