AI Ghost Citations Raise Trust Concerns in Research

AI-generated fake references are becoming a growing concern in academic publishing.

These false references are often called ghost citations. They look like normal academic sources, with author names, article titles, journals and publication years. The problem is that the cited paper does not exist.

The issue has become more urgent as researchers, students and editors use generative AI tools for writing, literature searches and reference management.

What are ghost citations?

A ghost citation is a reference to a non-existent source.

It may look convincing at first glance. It may include a real-looking title, a known journal name or the name of an actual researcher. But when someone checks the reference, the article cannot be found.

This is different from a normal citation mistake. A typo in a page number or year can often be corrected. A ghost citation points to research that was never published.

That matters because citations are the foundation of academic trust. They allow readers to check whether a claim is supported by previous research.

If the source does not exist, the claim may look stronger than it really is.

Why AI makes the problem worse

Large language models can produce text that sounds fluent and credible, but they can also invent facts. The same can happen with academic references.

A researcher may ask an AI tool to suggest recent studies, improve a literature review or format a bibliography. If the output is not checked carefully, fake references can enter a manuscript.

A 2026 preprint called GhostCite tested 13 large language models and found that all of them produced hallucinated citations at least some of the time. The reported hallucination rates ranged from 14.23% to 94.93%, depending on the model and research domain. The same study analysed 2.2 million citations from 56,381 papers in AI, machine learning and security venues, and found that 1.07% of papers contained invalid or fabricated citations.

Another 2026 preprint analysed 111 million references across 2.5 million papers in arXiv, bioRxiv, SSRN and PubMed Central. The researchers estimated 146,932 hallucinated citations in 2025 alone.

These are preprint findings, so they should be read carefully. But the direction is clear: fabricated references are no longer only an isolated embarrassment. They are becoming a measurable risk in research workflows.

Why this matters

A fake reference can do more damage than it first appears.

If a paper cites non-existent research, readers may assume the claim has support. Reviewers may miss the problem. Other authors may later copy the false reference into new work.

That can pollute the scientific record.

The risk is especially serious in fields where research affects real decisions, such as healthcare, climate, public policy, law, education and technology.

In medical research, a fake source can make an argument look more evidence-based than it is. In policy research, it can make weak claims look more settled. In AI research, it can inflate the credibility of technical claims that have not actually been proven.

Ghost citations also create extra work for editors and reviewers. Peer review already depends on limited time. If reviewers do not check references, fake citations can pass through the system.

Human responsibility remains central

The issue is not that AI can never be used in research.

AI tools can help with language editing, summarising, formatting and organising notes. The problem begins when AI output is treated as verified knowledge.

Major publisher policies already place responsibility on human authors. Elsevier’s generative AI policy says authors may use AI tools to help organise literature and suggest sources, but they must carefully review and verify all AI-generated output. It also warns that fabricated references may lead to rejection of a manuscript.

Publication ethics guidance makes the same point: AI cannot take responsibility for a paper. Authors, editors and reviewers remain responsible for accuracy.

That means a simple rule should apply: no reference should enter a paper unless a human has checked that it exists and supports the claim being made.

What universities and journals should do

The solution is not only telling individual researchers to be careful.

Universities and journals need clearer systems for AI-assisted research. Authors should disclose meaningful AI use. Journals should strengthen reference checks. Reviewers should be supported with better tools. Research groups should train students and staff to verify sources before submitting work.

Libraries and research support teams may also become more important. They know how to search databases, check DOIs and identify unreliable sources.

The challenge is speed. AI tools can generate a bibliography in seconds. Proper verification takes time.

But that time is part of research quality.

A warning for the AI era

Ghost citations show one of the clearest dangers of using AI in knowledge work.

The output can look professional even when it is wrong.

That makes the problem harder to spot. A fake citation does not always look suspicious. It may be formatted perfectly. It may use the right academic style. It may even include real author names.

But if the paper does not exist, the reference is worthless.

Academic publishing depends on traceable evidence. AI can help researchers work faster, but speed cannot replace verification.

The message for universities, journals and researchers is direct: AI can support research, but it must not be allowed to invent its evidence.

Sources: GhostCite preprint on citation validity in the age of large language models, preprint on non-existent citations across large publication datasets, Elsevier generative AI policies for journals and COPE guidance on editors responding to suspected AI use.