“Frankenstein Citations”: How to Spot Partially Wrong AI References

Frankenstein citations are AI-generated references that look correct at a glance — every field is filled, formatting is proper — but the combination of elements never actually existed. To spot them, search the title in quotation marks on Google Scholar, resolve every DOI at doi.org and compare the retrieved paper’s title with the citation, and read the actual source to verify the claim you’re citing. If the paper is real but doesn’t contain the claim you attributed to it, you still have a Frankenstein citation.

What Are Frankenstein Citations?

A Frankenstein citation is a reference that AI builds from real pieces but assembles into a fake whole. The author name exists. The paper title exists. The journal name exists. The year exists. The DOI resolves. But the exact combination — the paper as cited — never existed in any database.

The term was coined by Joe Shockman of Grounded AI and first described in detail by researchers at Nature (Naddaf & Quill, 2026), who documented how AI systems combine real bibliographic elements into plausible but fabricated references. One documented case involved researcher Guillaume Cabanac, who discovered his name attached to 23 papers he never wrote — with titles like “The Evolution of the Internet in the 19th Century,” which he never authored or co-authored.

For students, the danger is that Frankenstein citations pass the casual check. A reference that looks properly formatted, contains a real author name, resolves to a real DOI, and sits in a plausible journal — it doesn’t scream “fake.” That’s precisely what makes Frankenstein citations more insidious than fully hallucinated ones.

This article covers the middle ground: references that look correct enough to escape suspicion but contain errors that only surface under scrutiny.

• Frankenstein citations use real bibliographic elements assembled into combinations that never existed
• They are harder to detect than fully hallucinated references because they pass casual checks
• Three failure modes exist: Chimera (nonexistent combination), Distorted (wrong metadata), and Compound Deception (real paper, fabricated claim)
• Every detection method only catches specific failure types — no single method catches all Frankenstein citations
• The five-step verification checklist includes claim alignment verification (Step 5), which catches compound deception that DOI checks alone miss
• The Escalation rate has increased from 1-in-2,828 (2023) to 1-in-277 (early 2026)
• Using AI-generated citations without verification may constitute academic misconduct under the recklessness standard
• If you’ve already submitted a paper with Frankenstein citations, verify all references, correct errors, and disclose proactively

The Three Failure Modes

Detection requires knowing what to look for. Research identifies three distinct failure modes, each requiring a different verification method.

Chimera: Real Parts, Nonexistent Combination

A Chimera citation takes multiple real papers, real authors, and real journal names and merges them into a single entry that never existed. The “title” looks like a plausible paper title. The author names are real scholars. The journal is real. But no such paper was ever published.

Detection: Search the cited title in quotation marks on Google Scholar or your library catalog. If the exact title yields no results, you likely have a Chimera citation.

Concrete example: An AI might generate:

Smith, J., & Doe, A. (2024). Cognitive load and working memory in STEM education. Journal of Educational Psychology, 45(3), 112-128.

This looks correct. But “Cognitive load and working memory in STEM education” as a joint paper by these specific authors in this specific journal volume does not exist. Both authors exist in the field, the journal is real, the formatting is standard — but the paper is fabricated.

Distorted: Real Paper, Wrong Metadata

A Distorted citation points to a real paper but gets some metadata wrong: wrong volume number, wrong page range, wrong publication year, or a DOI that resolves to an entirely different paper. Unlike Chimera citations, a tool checking only for paper existence may flag this as “verified” — even though the specific bibliographic details are wrong.

Detection: Resolve every DOI at doi.org and compare the retrieved paper’s actual title and page numbers with the citation. Walters & Wilder (2023) found that 43% of citations generated by GPT-3.5 and 24% of those generated by GPT-4 that pointed to “real” papers still contained substantive metadata errors.

Concrete example: A citation might include a valid DOI but the DOI resolves to a paper with the same authors but a completely different title — or it resolves to a paper published in a different year than cited. The DOI works. The authors are correct. But you’re citing the wrong version.

Compound Deception: Real Paper, Fabricated Claim

This is the most deceptive failure mode. You locate a real paper. You verify the DOI. The authors are correct. The journal is real. You read the cited section… and the paper does not actually say what you’re citing it for. AI can cite a genuine source while attributing a fabricated claim or conclusion to it.

Detection: Read the actual source. Do not assume that a verified paper supports the claim it’s being used for. This step — claim alignment verification — catches compound deception that the first two detection methods miss entirely.

Concrete example: The AI cites a paper on student motivation that discusses growth mindset interventions. But the claim you’re writing — “student motivation is primarily driven by external rewards” — comes from an entirely different paper. The AI cited a real paper but attributed a claim from another paper to it. Ansari et al.’s (NeurIPS) analysis found that 63% of hallucinated citations exhibited this kind of semantic hallucination (real paper, fabricated claim), and 29% exhibited identifier hijacking (the DOI resolves but to a different paper than the one intended).

Why Frankenstein Citations Are the Most Dangerous Type

Full hallucinations — papers that never existed with any element correct — are easier to catch. A title search in quotation marks will fail. A DOI will not resolve. But Frankenstein citations pass every surface-level check.

The problem compounds when you consider the verification tools students typically use. The most common is Google Scholar’s basic search, which often returns “close enough” results when you don’t use exact quotation marks. A partially correct title may appear in related paper lists, reinforcing the illusion of correctness through a “verification by familiarity” effect: if the paper looks like it could exist, you don’t investigate further.

The Ansari et al. (NeurIPS 2025) analysis identified this as a systemic problem. Even top-tier peer-reviewed papers failed to catch these citations: GPTZero’s scan of 4,841 NeurIPS 2025 submissions found 100 hallucinated citations across 51 papers that passed peer review. The submissions were heavily AI-assisted, and reviewers — facing a 220% increase in submission volume (9,467 papers in 2020 to 21,575 in 2025) — had zero incentive to verify citations that bring no reward and carry no risk of being caught by automated tools.

A Five-Step Detection Checklist for AI-Generated Citations

Every AI-generated citation should pass through this verification sequence. Each step catches a different failure mode.

Steps 1–4 verify the existence of the cited paper. Step 5 verifies the use of the cited paper. Most existing verification guides stop at Step 4. The Ansari et al. data shows that Step 4 is insufficient: 43% of “real” citations still contained errors detectable only at Step 5.

The Scale Is Escalating

The problem is not static. The Lancet audit data shows escalating fabrication rates:

• 2023: 1 in 2,828 papers contained fabricated citations
• 2025: 1 in 458 papers
• Early 2026: 1 in 277 papers

Grounded AI’s analysis of 110,000+ scientific publications from 2025 found at least one AI-generated invalid reference in each. The rate is accelerating as AI writing tools become more sophisticated at producing plausible-looking references.

What Happens When You Submit a Frankenstein Citation?

The consequences depend on whether the error was intentional or accidental.

Under the Resnik & Hosseini (2026) framework published in Accountability in Research, using AI-generated citations without verification may constitute research misconduct under the recklessness standard. The recklessness standard applies when a researcher shows “indifference to a known risk of fabrication.” If you know AI can invent references, you know you must check them. Failure to check is not “an honest mistake” — it is reckless.

Academic integrity policies increasingly treat fabricated references as misconduct in and of themselves. The TEQSA (2026) Australian regulator explicitly lists fabricated references as a form of academic misconduct, regardless of whether the student intended to deceive. The Watson (2024) University of Mississippi study found that over 90% of student submissions with hallucinated citations were never verified before submission.

Professional consequences extend beyond academia. Legal precedent (e.g., Mata v. Avianca, the case in which attorneys cited a non-existent Supreme Court case) has produced sanctions up to $5,000 per fabricated citation. In professional writing, fabricated citations can void publications, trigger journal retraction notices, and permanently damage a researcher’s reputation.

How to Verify AI-Generated Citations (Practical Guide)

The most effective verification workflow for students:

Step 1: When AI provides a reference, copy the title and search it on Google Scholar with quotation marks around the entire title. If Google Scholar returns zero results, the paper likely does not exist.

Step 2: Resolve every DOI at doi.org (not Google Scholar). Read the retrieved paper’s title. Compare it with the cited title. If they don’t match, the citation has a metadata error.

Step 3: Read the actual paper for any claim you’re attributing to it. Even if the paper is real, even if the DOI resolves correctly — verify that the specific claim, data point, or conclusion you’re citing actually appears in the paper.

Step 4: Check the paper’s author list and publication date against the citation. Authors may share names. Publications may have different volume numbers across databases. Cross-reference with the journal’s official website.

Detection Decision Tree: Which Method Catches Which Failure Type

Not all detection methods catch all types of Frankenstein citations. Using the wrong verification tool for the wrong failure type is why these citations survive scrutiny. This decision tree maps each failure type to its specific detection method and red flags.

What to Do If You’ve Already Submitted a Frankenstein Citation

If you discover a Frankenstein citation in a paper you’ve already submitted:

1. Assess the scope. How many citations are suspect? Are they in a literature review, methodology section, or supporting bibliography? Citations in the core argument carry more risk than peripheral references.
2. Verify everything. Run all citations through the five-step checklist above. Document which citations are verified and which fail verification.
3. Disclose proactively. If the paper has not yet been finalized (e.g., before submission to a journal, before a thesis defense), correct the citations before submission. Many institutions view proactive correction more favorably than discovery through external audit.
4. If discovered post-submission, contact your advisor or journal editor with a written explanation of the error and the corrected citations. The Resnik & Hosseini framework treats unverified AI citations as recklessness only when there is knowledge of the risk and indifference to it. Demonstrating correction and prevention measures can mitigate this finding.