Why does ChatGPT create fake references instead of saying it doesn't know?

ChatGPT is trained to produce fluent, plausible-sounding language using probabilistic next-token prediction. It does not perform a live search against Crossref or PubMed unless specifically integrated with an external search tool. It generates what a citation typically 'looks like' rather than retrieving verified facts.

Can Turnitin detect if a citation was made up by ChatGPT?

Turnitin detects text similarity and AI-generated writing style, but it does NOT verify citation truthfulness against scholarly registries. A fabricated paper title can pass Turnitin's plagiarism report with 0% similarity because the text is original, even though the citation is fraudulent.

How can I tell if a reference produced by an AI is real?

Paste the DOI into https://doi.org/ and search the exact article title in Crossref (search.crossref.org) or PubMed. If no matching record exists in official indexing registries, the citation is an AI hallucination.

AI Hallucination Alert Reading time: 6 min

Can ChatGPT Generate Fake Citations? Understanding AI Hallucinations in Research

By CitePure Editorial Lab • November 14, 2026 • Technical Analysis

💡 Direct Answer: Yes, ChatGPT Routinely Fabricates Citations

Yes. Unless explicitly connected to an external scientific database, Large Language Models (LLMs) like ChatGPT, Claude, and Gemini frequently invent non-existent references. Studies show that between 30% and 55% of references generated by generic LLMs contain phantom DOIs, mismatched authors, or entirely fabricated article titles.

This occurs because LLMs are generative token predictors—not search engines or knowledge databases. They generate citations that statistically mimic scholarly formats rather than retrieving actual publications.

1. The Mechanical Cause: Why Do LLMs Hallucinate Citations?

To understand why an AI model invents citations, you have to look under the hood of Transformer architectures. When you ask ChatGPT to "provide 5 APA 7th edition sources supporting the impact of microplastics on cognitive function," the model does not search PubMed, Web of Science, or Scopus.

Instead, it generates words based on statistical probability (Next-Token Prediction):

It knows that in this discipline, researchers with names like "Smith," "Chen," or "Müller" frequently publish papers.
It knows words like "Environmental Health Perspectives" or "Journal of Hazardous Materials" are common journal names.
It knows the structure of an APA citation: Author, A. (Year). Title. Journal, Volume(Issue), Pages. DOI.

By blending these probable elements together, the model outputs a citation that sounds completely authentic to the human reader, but has never existed in print.

2. The 3 Types of AI Citation Hallucinations

In CitePure's automated audits of over 50,000 AI-assisted student drafts, citation hallucinations consistently fall into three distinct archetypes:

1. The Phantom Reference (100% Synthetic)

The author name is generic, the title is completely synthesized, the volume and pages are random numbers, and the DOI (e.g., https://doi.org/10.1016/j.envres.2023.115890) returns a hard 404 error on the International DOI resolver.

Risk Level: Catastrophic. Easily caught by any reviewer or professor who clicks the DOI.

2. The Chimeric Citation (Real Author + Real Journal + Fake Title)

The AI identifies a real, prominent researcher in the field (e.g., Yoshua Bengio) and pairs them with a real journal they frequently publish in (e.g., Nature Machine Intelligence). However, the specific paper title never existed. The AI creates a "chimera" out of real components.

Risk Level: Extreme. Reviewers often assume the citation is real because the author and journal are prestigious.

3. The Impersonated DOI (Real DOI, Wrong Subject)

The model provides a working DOI that successfully resolves to a publisher landing page. However, when clicked, the DOI belongs to an article about pediatric cardiology, while the citing paper is discussing macroeconomic trade policy.

Risk Level: Fatal. Proves to academic tribunals that the author never read the cited paper.

3. Why Plagiarism Checkers Won't Protect You

Tool Type	What It Checks	Can It Detect a Fake Citation?
Turnitin / iThenticate	Text similarity against student papers and web pages	NO (AI text is often 0% plagiarized)
Grammarly	Spelling, grammar, and generic style conformity	NO (Considers fake citations grammatically valid)
Zotero / Mendeley	Formats saved library entries into reference lists	PARTIAL (Only formats what you manually save)
CitePure Reference Verifier	Deterministic API lookup against Crossref, OpenAlex & PubMed	YES (Validates existence of paper, authors & DOI)

4. How to Safely Use AI for Academic Literature Search

If you choose to use AI during the preliminary stages of research, follow these non-negotiable safety rules:

Never copy-paste AI references directly into your paper: Treat every AI-generated citation as a hallucination until proven otherwise.
Verify the source document: Open Google Scholar, PubMed, or Crossref, paste the title, and download the actual PDF. If you cannot obtain the PDF, delete the citation.
Check the claim against the text: Even when a paper is real, LLMs frequently misrepresent the paper's findings, claiming a study proved X when it actually concluded Y.