Tags: #academic-integrity, #AI-hallucination, #AI-writing, #citation-verification

What Are Frankenstein Citations?

Frankenstein citations are AI-generated references that assemble fragments of real scholarly work into a citation that looks complete but describes a paper that never existed. The term was coined by Joe Shockman, co-founder of Grounded AI, in analysis covered by Nature in April 2026. He likened the citations to the fictional Frankenstein monster: assembled from pieces of real things, but recombined into something that doesn’t actually exist. See Nature’s analysis for the full coverage.

  • Frankenstein citations are AI-stitched references that combine real fragments (author names, journal titles) with fabricated ones (wrong titles, invented volumes) into citations that look correct but describe papers that never existed
  • Every hallucinated citation in the NeurIPS 2025 audit exhibited compound failure modes — real and fake elements layered together to evade detection
  • A working DOI doesn’t prove a citation is correct; a DOI that resolves to a different paper is just as fabricated as a broken DOI
  • The five-step verification checklist below catches Frankenstein citations that automated tools and peer reviewers miss

Here’s exactly what a Frankenstein citation looks like:

  • The author name is a real researcher with real publications
  • The journal title is a real, legitimate academic journal
  • The paper title sounds exactly like something they would have written
  • The DOI, year, and volume are formatted correctly

But the actual paper does not exist. No paper with that title was ever published in that journal by that author. The AI simply predicted the most plausible-looking combination of bibliographic fragments.

This isn’t a “maybe” problem. It’s a documented, peer-reviewed reality that affects student papers, peer-reviewed conference proceedings, government reports, and professional consulting outputs.

The most reliable indicator of a Frankenstein citation is what researchers call the “uncanny valley” tell: the citation is structurally perfect, all the fields are present, all the formatting is correct, but the individual pieces have never actually belonged together. Every single element checks out when you look at it separately. Only when you verify the combination do you discover the paper was invented.

How It’s Different from “Vibe Citing”

The term “vibe citing” — introduced by GPTZero — describes the same phenomenon from the LLM’s perspective: citations that sound plausible enough to survive a reviewer’s quick glance. While “Frankenstein citation” describes the structure of the reference itself (real parts, fake combination), “vibe citing” describes the mechanism (the model predicts a plausible sequence and the reviewer trusts the vibe rather than verifying). GPTZero provides detailed detection methods for spotting fake citations.

How Frankenstein Citations Are Built

An LLM doesn’t look up a database when it generates a citation. It predicts the next plausible string of text from patterns it has seen in millions of published references. When you ask for “three peer-reviewed studies on X,” the model produces:

LitSynth explains the underlying mechanism: LLMs are not databases. They’re next-token predictors trained to produce plausible text. Citations are generated, not retrieved. See LitSynth on why LLMs invent citations.

When you ask for “three peer-reviewed studies on X,” the model produces:

A name from that field + a journal that publishes in that field + a plausible title + a DOI in the right format.

Every individual token is statistically probable. The combined reference reads perfectly. But the underlying paper does not exist.

The Five Failure Modes

A 2026 study by Samar Ansari et al. analyzed 100 AI-generated hallucinated citations in papers accepted by NeurIPS 2025 — one of the world’s most prestigious AI conferences — and developed a five-category taxonomy that explains how these citations fail. The full paper is available on arXiv.

Failure Mode What It Means Real-World Impact
Total Fabrication (66%) The citation has zero overlap with any real scholarly work. Authors, title, venue, and identifiers are all invented. The AI creates a completely fictional paper from scratch.
Partial Attribute Corruption (27%) The citation blends real and fabricated elements — real authors attributed to papers they didn’t write, real paper titles with fabricated authors, correct venues with wrong publication years. A familiar name creates a false sense of legitimacy while bibliographic details are wrong.
Identifier Hijacking (4%) The citation uses a valid scholarly identifier (DOI or arXiv ID) that points to a real paper, but the authors, title, or other metadata don’t match the paper at that identifier. A reviewer who clicks the link sees a real paper and may assume the citation is correct.
Placeholder Hallucination (2%) The citation contains obvious generation failures such as template variables (“Firstname Lastname”) or incomplete identifiers. The model failed to complete its task. Easy to catch but indicates poor verification.
Semantic Hallucination (1%) The citation invents a plausible-sounding title that fits the research domain conceptually but doesn’t correspond to any real work. The title sounds professionally appropriate for the research context while corresponding to no actual publication.

Here’s the critical finding from Ansari’s analysis: every single hallucination (100%) exhibited compound failure modes. No hallucination was a simple, single-mode error. Real and fake elements were layered together to exploit multiple verification heuristics simultaneously. This is why Frankenstein citations are so dangerous.

Real-World Examples

Understanding Frankenstein citations is easiest with concrete examples. Here are three real cases that show how the AI-stitched reference works in practice.

Example 1: The Tversky & Kahneman Chimera

A chimera citation is assembled from real parts that never belonged together:

Tversky, A., & Kahneman, D. (1979). Heuristic anchoring under conditions of statistical uncertainty. Cognitive Psychology, 11(2), 168-192.

Tversky and Kahneman are real, foundational researchers. Cognitive Psychology is a real journal. 1979 is inside their active collaboration. Every part checks out individually.

But that exact paper with that exact title was never published.

The tell: a search for the exact title in Google Scholar returns nothing. The authors, journal, and year are real. The paper is not.

Example 2: The Bandura Distorted Citation

A distorted citation is a real paper underneath, but with details mangled. Bandura’s famous paper exists:

Bandura, A. (1977). Self-efficacy: Towards a unifying theory of behavioral change. Psychological Review, 84(2), 191-215.

A distorted version might read:

Bandura, A. (1978). Self-efficacy: Towards a unifying theory of behavioral change. Psychological Review, 84(3), 191-215.

The year is off by one. The issue number is shifted. A casual check might miss it entirely. But the claim you attribute to it may not be supported by the actual text.

The tell: resolve the DOI at doi.org and compare every field — title, authors, year, volume, issue, pages — against the citation. If any field disagrees, the citation is distorted.

Example 3: Cabanac’s “Tortured Phrases”

In 2021, computer scientist Guillaume Cabanac and colleagues noticed strange phrases in journal articles — “counterfeit consciousness,” “profound neural organization,” “colossal information” — that academics never use naturally. These were AI-generated paper titles, invented by LLMs trained on scientific literature.

Nature reported on Cabanac’s work, documenting how AI models generate paper titles that sound authoritative but correspond to no published work. The “tortured phrases” tell is linguistic: real researchers use standard terminology; AI-generated titles use awkward constructions that no human would choose. See Nature’s coverage of Cabanac’s study.

The Scale of the Problem

Frankenstein citations aren’t limited to student papers. They have surfaced in peer-reviewed conference proceedings, government reports, and professional consulting outputs.

NeurIPS 2025 Audit

A 2026 audit of papers accepted at NeurIPS found 100 hallucinated citations across 53 papers that were missed by 3-5 expert reviewers per paper. These aren’t typographical errors — they are citations to sources that do not exist, with invented authors, fabricated titles, and false publication details. GPTZero’s full report documents every hallucinated citation across the accepted papers.

“The appearance of AI-generated hallucinated citations in peer-reviewed literature represents a fundamental challenge to the integrity of scientific discourse.” — Ansari et al., arXiv 2026

arXiv One-Year Ban

In May 2026, Thomas Dietterich — chair of arXiv’s computer science section — announced a new rule: authors who submit papers with unchecked LLM-generated content, including hallucinated citations, face a one-year ban from the platform. See GPTZero’s coverage for the full announcement and context.

“By signing your name as an author of a paper, each author takes full responsibility for all its contents, irrespective of how the contents were generated.” — Thomas G. Dietterich, arXiv, May 14, 2026

Watson (2024) Student Study

A 2024 study by Alex P. Watson at the University of Mississippi examined AI-generated citations submitted by freshmen. The findings are important for understanding the scope of the problem. The study is published in The Serials Librarian (Vol. 85, Issue 5-6, 2024, pages 172-180). See Google Scholar for the full article.

  • Most suspect citations were not random nonsense: they mixed real fragments (author names, journal titles, publication years) with invented ones (volume numbers, page ranges, article titles)
  • 93% of citations (43 of 46) mixed real components with fabricated details
  • Real authors appeared in 39% of citations; real titles in 22%; real journals in 89%; but real URLs/DOIs appeared in 0% of hallucinated citations
  • Most student submissions with hallucinated citations were never verified before submission
  • Submitting fabricated sources, even unintentionally, frequently triggers academic integrity reviews

The Frankenstein Verification Checklist

This is the primary tool for verifying any AI-generated citation. Run it for every reference before you submit.

Step Action Tool What to Look For Red Flag
1 Exact-title search Google Scholar (quotation marks), Crossref Does the paper appear with exact title and same authors? Zero results, or only unrelated papers
2 DOI resolution doi.org Does the DOI resolve to a real article? “DOI Not Found” error or resolves to a different paper
3 Metadata cross-check DOI resolve + CrossRef or PubMed Do title, authors, year, journal all match? Any field disagrees with the official record
4 Author verification ORCID, Google Scholar profile, institutional listing Does the named author have this paper in their publication list? Author is real but never published this title
5 Claim alignment check Read the actual source Does the paper support the specific claim you’re citing? The paper exists but doesn’t say what you claim

Bottom line: If any step fails, the citation is fabricated, distorted, or compound-deceptive. Do not cite it.

Step 1: Exact-Title Search

Copy the entire title into quotation marks. Paste it into Google Scholar or search.crossref.org. If the paper was actually published, it will appear as the first result. If the AI made up the title, the search will return nothing or unrelated papers.

What happens in practice is straightforward. Silence is the strongest signal of an invented reference. A real, published, indexed paper with a specific title almost always shows up in Google Scholar.

Step 2: DOI Resolution

Paste the DOI after https://doi.org/ in your browser. The resolver redirects you directly to the publisher’s page. What happens next matters: read the title, the authors, and the publication year on the landing page.

Don’t just confirm that the DOI resolves. Read the landing page. A DOI can resolve to a real paper, and the real paper is completely unrelated to the citation you were given. This is the hidden danger of Frankenstein citations.

Compare every field on the resolved page against your citation: title, authors, journal, year, volume, issue, pages. If any field disagrees, the citation is wrong.

Step 3: Metadata Cross-Check

Open the paper’s metadata on Crossref or PubMed. Compare the official record against your citation line by line. You can use Crossref’s public API directly in your browser:

https://api.crossref.org/works/[DOI]

Replace [DOI] with the actual identifier. The API returns JSON with full metadata. This is faster than web searches and useful when checking dozens of references at once.

Step 4: Author Verification

Check the author’s own publication list. Most researchers list every paper they’ve written on their ORCID profile, Google Scholar page, or university directory. If the paper you’re citing isn’t on that list, it probably wasn’t written by that author.

Pro tip: If you’re using Zotero, install the free Retraction Watch Integration plugin. It automatically flags retracted items in your library using privacy-preserving partial matching.

Step 5: Claim Alignment Check

Confirm the paper supports the specific claim you’re citing. This is where most students cut corners. They resolve the DOI, confirm the paper exists, and then proceed to quote it without reading the relevant section. They assume the AI described the paper accurately. That assumption is often wrong.

The way to catch compound deception is simple but time-consuming: read the section you’re citing. Don’t skim. Don’t rely on the abstract. Go to the actual paragraph, read the surrounding context, and confirm the specific claim is there.

The Frankenstein Anatomy: Real vs. Fake Breakdown

Use this comparison table to quickly diagnose which type of Frankenstein citation you’re dealing with.

Fragment What’s Real What’s Fake Verification Step
Author name Real researcher with real publications Never wrote this specific paper Check ORCID or institutional listing — does this author have this paper?
Journal title Real, legitimate academic journal Never published this article Search the journal’s archive for the volume, issue, and page numbers you cite
Paper title — Invented entirely Search the exact title in quotation marks on Google Scholar
DOI May be real (borrowed from another paper) May be fabricated (never registered) Resolve at doi.org — compare the landing page against your citation
Year / Volume / Pages — Invented or transposed Cross-check against CrossRef metadata or the journal’s own page

This table shows why Frankenstein citations are so dangerous: every individual fragment passes a quick check. Only when you verify the combination of the specific author writing the specific paper in the specific journal with those specific details do you discover the reference was invented.

What to Do When You Find a Frankenstein Citation

So you’ve run the checklist, and one of your AI-generated citations failed. Maybe the title returned zero results. Maybe the DOI resolved to a completely different paper. Now what?

Replace the fake citation with a real one that supports the same claim.

Don’t try to “fix” the DOI or edit the year. If the paper doesn’t exist, you need to find a real source for the point you were trying to make.

Search Strategies for Replacement

  • Google Scholar: Search the key concepts of your claim, not the fake title. Sort by relevance, scan the top results, and read the abstracts to confirm a paper genuinely supports your point.
  • Subject databases: PubMed for medicine and life sciences, CrossRef and Semantic Scholar for broad coverage, plus discipline-specific indexes.
  • Your library: University libraries give you licensed access to paywalled databases and, crucially, a research librarian whose job is helping you find the right source fast.

Read enough of each candidate to confirm it actually says what you’re claiming. This is the scholarship the AI skipped, and the reason the final work ends up yours instead of a model’s guess.

Once you’ve replaced the citation, check the rest of your references. Fake citations tend to cluster because the conditions that produced one were in force for the whole session: the model guessing instead of retrieving.

Why This Matters More Than Ever

Frankenstein citations aren’t a quirk of AI writing tools. They’re a predictable byproduct of how large language models generate text: they predict the next plausible sequence, not verify whether the source exists. The result is a citation that looks perfect to anyone who glances at it: real author name, real journal title, correct formatting, but describes a paper that was never written.

The Institutional Response

The problem has triggered concrete institutional responses. arXiv’s one-year ban (May 2026) established the first major policy: authors are fully responsible for the contents of their work, and unchecked AI-generated citations are grounds for platform exclusion. NeurIPS itself has an LLM policy that considers hallucinated citations grounds for paper rejection or revocation.

Resnik & Hosseini (2026) argue in Accountability in Research that hallucinated citations may constitute research misconduct when citations function as data in scholarly papers. Their analysis shows that when a researcher uses a GenAI tool without checking its output for veracity, they are acting recklessly because they’ve shown indifference to the risk of data fabrication. The full paper is available on PubMed Central.

“To act recklessly means to propose, perform, or review research, or report research results, with indifference to a known risk of fabrication.” — Resnik and Hosseini, 2026

The Practical Bottom Line

For students and researchers: “The AI gave it to me” is not a defense. Every university treats citation accuracy as your responsibility, not the AI’s. Verifying what you cite is not optional.

AI assistance in research is normal and acceptable. AI-suggested citations are not. Use AI for structuring, drafting, and brainstorming — our guide on the best AI tools for academic writing compares the main options. Do not let AI generate references.

Here’s what I recommend starting with: run every AI-sourced citation through the five-step checklist above. It takes about two minutes per citation. That’s less than ten minutes for a paper with five references. The cost of verification is trivial. The cost of citing a fabricated source ranges from a lowered grade to an academic integrity hearing.

Don’t trust the AI to verify its own work. Don’t trust a “working” DOI to prove a paper is real. Don’t trust the vibe.

Conclusion

Verification is a habit, not a choice. Frankenstein citations are real. They’re documented in peer-reviewed studies. They’ve surfaced at the world’s most prestigious AI conferences. They’re in student papers.

The five-step checklist above is the tool. Use it for every AI-generated citation. Check the title, resolve the DOI, cross-check the metadata, verify the author, read the passage behind the claim. If any step fails, replace the citation.

And remember: using AI tools for research is fine — but verifying every citation is not. That’s what makes your paper yours and not a model’s guess.

If you want to dig deeper into citation verification, check our guide on managing citations with Zotero, Mendeley & EndNote, or learn how to cite AI in academic writing with step-by-step formatting guidance. You can also read our guide on how to use AI for research without plagiarizing for responsible AI use strategies, or compare the best academic writing tools for your workflow.