The fastest way to fail a thesis defense

Did AI invent your sources? Check in 30 seconds.

ChatGPT and other AI tools generate references that look completely real, with plausible authors, a fitting journal, even a DOI, and point to papers that don't exist. Paste your reference list into the free quick-check and see which of your sources actually hold up.

Demo

Free bibliography check.

Paste your bibliography. Acurio instantly checks every entry against Crossref & OpenAlex and flags retracted papers. Try it right here.

Paste your bibliography. Acurio instantly checks every entry against Crossref & OpenAlex and flags retracted papers.

Free, no account. Checks each reference for existence and retraction against Crossref, OpenAlex, and DOAJ.

What is a hallucinated citation?

A hallucinated citation is a reference an AI model invented. It reads like the real thing: real-sounding author names, a journal that exists, a year that fits, a DOI in the right format. It comes in two flavors, and the second one is harder to catch.

The source doesn't exist

The paper was never written. No database has it, the DOI resolves nowhere or to something unrelated. This is the classic fabrication: detectable if you look every reference up, which almost nobody does for a full bibliography.

The source exists but says something else

The paper is real, the claim attached to it is not. The model paired a real reference with a statement its authors never made. No existence check catches this. You have to read the source against the claim.

Why language models fabricate references

Language models generate the most plausible next words, not verified facts. A citation is just text with a very predictable shape (Author, Year, Title, Journal), so the model produces something with exactly that shape. Whether the paper exists never enters the process.

Unless a tool is explicitly wired to a real bibliographic database, it has no way to look anything up. It completes the pattern. That's also why fabricated references are so convincing: they're optimized to look right, not to be right.

Newer models fabricate less, but not zero. And the errors that remain look more plausible, not less.

The data

How often does it happen?

Peer-reviewed studies have measured fabrication rates directly. The numbers are not edge-case numbers.

19.9%
of GPT-4o references were entirely fabricated

In a 2025 JMIR Mental Health study, one in five references GPT-4o generated for mental-health literature reviews didn't exist at all, and 45% of the real ones contained bibliographic errors. Roughly two-thirds were fake or wrong.

JMIR Mental Health, 2025 (Deakin University)
55% → 18%
fabricated citations, GPT-3.5 vs GPT-4

A 2023 study in Scientific Reports found 55% of GPT-3.5's citations were fabricated, against 18% for GPT-4. Better models fabricate less. None fabricate nothing.

Scientific Reports (Nature), 2023
DOI ≠ real
most fake DOIs resolve to a real, unrelated paper

The same study found that most fabricated citations carrying a DOI pointed to a real but unrelated paper. A working DOI link is not proof the citation is right. You have to check what's behind it.

Scientific Reports (Nature), 2023

Fabrication rates vary by field, task, and model version; the studies above report the setups they measured. The direction is consistent everywhere: real, non-trivial rates.

Documented cases

This is already ending careers

Not hypotheticals. Courts sanction lawyers over invented cases; journals retract papers over invented references.

Mata v. Avianca, the landmark case

In June 2023, a federal judge in New York sanctioned two lawyers $5,000 after their brief cited six court cases that ChatGPT had invented. They were ordered to notify every judge falsely cited.

Mata v. Avianca, S.D.N.Y. 2023

$10,000 fine in California

In October 2025, a California appeals court fined an attorney $10,000 after finding 21 of 23 quotes in his opening brief were AI-fabricated. Reported as the largest such fine by a California court at the time.

The Daily Record, Oct 2025

Over a thousand documented court cases

A live academic tracker documents court decisions worldwide where parties submitted AI-hallucinated material. It passed a thousand entries in 2025 and keeps growing.

AI Hallucination Cases database (D. Charlotin)

And it's not just lawyers

Fabricated citations passed NeurIPS peer review

A 2026 analysis found roughly 1% of papers accepted at NeurIPS 2025, a top machine-learning conference, contained fabricated citations that reviewers never caught.

arXiv:2602.05930

“Polluting the scientific literature”

Nature's news desk reported in 2026 on hallucinated citations spreading through published papers: errors that get cited onward by researchers who never open the original.

Nature news, 2026

Potentially research misconduct

A 2026 paper in Accountability in Research argues that submitting fabricated citations can constitute research misconduct, the same category as data fabrication.

Accountability in Research, 2026

How to check if a citation is real

By hand (fine for three references)
Search the DOI on Crossref

If the reference has a DOI, look it up on crossref.org. No match, or a match to a different paper: red flag.

Search the title in Google Scholar

Exact-match the title in quotes. A real paper shows up with its authors and venue. Check the year and author names actually match your reference.

Open the source and find the claim

Existence isn't enough. Read the passage you're citing: does the paper actually say what your text claims it says?

Automatically, for a whole reference list

Paste your bibliography into the free quick-check: every reference is checked for existence and retraction in seconds. For the second failure mode (real source, wrong claim), Acurio's full analysis compares every claim in your thesis against the actual text of your uploaded sources.

Run the free quick-check

Frequently asked questions

What is a hallucinated citation?
A reference that looks completely real, with plausible authors, journal, even a DOI, but points to a source that doesn't exist, or to a real paper that doesn't say what you cited it for.
Why does ChatGPT make up references?
Language models generate the most plausible-looking text, not verified facts. Citation details are easy to blend into something that reads correctly but isn't real, unless the tool is connected to a real bibliographic database.
How common are fake AI citations?
In a 2025 JMIR Mental Health study, about one in five references GPT-4o generated for literature reviews were entirely fabricated, and nearly two-thirds were fake or contained errors.
How do I know if a citation is real?
Search the DOI on Crossref, look the title up in Google Scholar, and open the source to confirm it supports your claim. A working DOI isn't enough; fabricated citations often carry a DOI that opens a real but unrelated paper.
Can Turnitin catch AI-hallucinated citations?
No. A fabricated citation is original text, so it passes a similarity check clean. Turnitin checks for plagiarism and AI-generated writing; verifying that sources exist and support their claims is a different check.
Is using an AI-generated citation academic misconduct?
It can be. A 2026 paper in Accountability in Research argues fabricated citations may constitute research misconduct, and journals have retracted papers over them. Intent doesn't always matter: you're responsible for your bibliography.

Check your references before your examiner does.

Paste your reference list: existence and retraction checked free, no account. Full claim-by-claim verification when you're ready.