A politics student I know got a 14 in a 3,000-word essay. Not fourteen out of one hundred because the argument was bad. Fourteen because three of her seven citations did not exist. She had asked ChatGPT for "a few more recent sources on populism" and it obliged with three perfectly formatted, perfectly plausible, perfectly fictional references. The authors were real. The journals were real. The article titles, DOIs and page ranges were hallucinated. Her tutor spotted it in under a minute and referred it to the academic misconduct panel.

This is now a routine category of failure. Generative AI does not know the difference between a real paper and a paper-shaped sentence, and it will give you the paper-shaped sentence with total confidence. If any of your sources came from a chat with an AI, or even from a search result that was auto-summarised by an AI, you need to check they are real before you submit. Here is how.

The short answer: how to check if citations are real

To check if a citation is real, look up the DOI on doi.org, search the exact article title on Google Scholar and the journal's own website, and cross-reference the author against their institutional page. Any citation that fails two of these three checks is almost certainly fabricated. For a full reference list, paste it into a verifier that queries CrossRef, which is the metadata registry every legitimate academic paper is indexed in, so you catch fakes in seconds rather than minutes.

The rest of this guide explains why AI fabricates citations, what a fake reference looks like up close, the manual checks that actually work, and the faster automated route for when you have thirty sources and forty minutes.

Why AI invents citations in the first place

Large language models do not retrieve papers. They generate text token by token based on statistical patterns in their training data. When you ask for "three recent sources on climate adaptation in coastal cities", the model is not searching a database. It is predicting what a plausible citation to that topic would look like, based on the shapes of millions of real citations it has seen.

The result is a reference that looks correct in every surface-level way:

Author names are real academics who work on the topic.
The journal is a real journal in the field.
The year is recent and plausible.
The volume and issue numbers are internally consistent.
The DOI has the right format (10.xxxx/xxxxx).

But the actual combination — this author, publishing this title, in this journal, in this year, with this DOI — does not exist. It is a statistical hallucination. The DOI will not resolve. The journal's archive will not contain the article. The author has never written a paper with that title.

This happens with every current general-purpose model, not just older ones. Retrieval-augmented models (the ones with browsing enabled) are much better but still occasionally hallucinate citations from web summaries or paywalled previews. If a citation entered your reference list through any AI tool, treat it as unverified until proven otherwise.

What a fabricated citation looks like

Here are three real patterns I have seen in student drafts, with the tells underlined in explanation.

Pattern 1: The real author, fake paper

Tufekci, Z. (2019). Algorithmic amplification and protest movements: Evidence from the Arab Spring decade. Journal of Communication, 69(4), 512-534. https://doi.org/10.1093/joc/jqz041

Zeynep Tufekci is a real scholar who has written extensively on protest movements and social media. The Journal of Communication is a real journal. The volume and issue structure is plausible. But she did not write that paper. The DOI resolves to a different article by a different author. This is the most dangerous fabrication because the name recognition makes it feel legitimate.

Pattern 2: The Frankenstein citation

Smith, J., & Patel, R. (2022). Rethinking digital literacy in higher education: A systematic review. Review of Educational Research, 92(3), 287-315.

Everything is real in isolation. Smith and Patel is the most generic author combination in English-language academia. Review of Educational Research is a real journal. The title is the kind of sentence a model generates constantly. But no such paper appears in any index. There is no DOI. When you search the exact title on Google Scholar, you get nothing.

Pattern 3: The dead-on-arrival DOI

Ahmed, S. (2023). Post-colonial readings of digital infrastructure. Media, Culture and Society, 45(2), 201-222. https://doi.org/10.1177/01634437221091324

The DOI looks fine, the right prefix, the right length. Paste it into doi.org and you get a 404 or an unrelated paper. This is the fastest fake to catch because DOIs are deterministic; either they resolve to exactly the right paper or they do not.

The three-check manual method

For any individual citation you are unsure about, run these three checks in order. If it fails two of three, it is almost certainly fabricated.

Check 1: Resolve the DOI

Go to doi.org and paste the DOI, or just append it to https://doi.org/. A real DOI redirects to the article's landing page on the publisher's site. The page title should match the citation title exactly, or be an obvious near-match accounting for sub-title differences.

If there is no DOI at all, that is not automatically a red flag (older articles sometimes predate DOIs), but it moves the citation to needing the other two checks done carefully. Every paper published in a major journal after about 2000 has a DOI.

Check 2: Search the exact title in quotes

On Google Scholar (scholar.google.com) or the journal's own search, paste the article title inside double quotes. This forces an exact-phrase match. A real article appears as the first result. A fabricated article returns zero results, or returns results for real but differently-titled papers by the same author.

If the exact-phrase search returns nothing on both Google Scholar and the journal's website, the article does not exist. Full stop.

Check 3: Check the author's publication list

Most academics have an institutional page, a Google Scholar profile, or an ORCID. Go to one of these and scan their publication list for the title in question. If the author is real and the paper is real, it will be listed. If the author is real but the paper is missing from their own CV, you are holding a fabrication with a real name attached.

This third check is what catches pattern 1 above, where the author is famous enough to make the fake feel credible. Famous scholars keep their publication lists up to date. If the paper is not on their own page, it does not exist.

CrossRef: the database every real citation is in

CrossRef is the non-profit metadata registry that publishers use to issue DOIs. Every academic paper with a DOI is registered with CrossRef, along with its title, authors, journal, year, and page numbers. There are more than 140 million records. Querying CrossRef directly is the most reliable fake-catcher that exists, because it bypasses search engines that sometimes index random web pages as if they were papers.

You can query CrossRef by DOI, title, or author-plus-year. A real citation returns an exact match. A fabricated one returns either nothing or a near-miss that does not match the citation you were given.

Rubrica's reference verifier is built around this. You paste your reference list, it parses each entry, queries CrossRef for all of them at once, and flags any that do not resolve. For a twenty-source bibliography this takes about fifteen seconds and catches every DOI-based fabrication. It is the single highest-value check if you have used any AI tool at any stage of your research.

The specific failure modes and how to catch them

Wrong year, right everything else

AI sometimes shifts the publication year by one or two to make a citation feel "more recent". The DOI will still resolve, but to the correct year, not the cited one. When you resolve the DOI and find the publisher page says 2019 but your citation says 2022, the citation is wrong. Fix the year, do not leave the fake year in your bibliography.

Real paper, wrong authors

A less common fabrication. The title resolves on Google Scholar, but the authors listed on the paper are different from those in your citation. This usually happens when a model conflates two papers on the same topic. Fix it by using the actual author list from the journal page.

Book chapters that do not exist

Edited volumes are a common fake-citation vector because the full chapter list is rarely online. The book is real, the editors are real, but the chapter titled "Chapter 7: Digital Dissent in the Global South" never appeared in the volume. Check the book's table of contents on the publisher site or Google Books preview. If the chapter is not in the ToC, it is a fake.

Preprints and grey literature

Preprints (arXiv, SSRN, OSF, PsyArXiv) are trickier because they can be real without a traditional DOI. Verify by searching the preprint title directly on the relevant server. If it is not there, it does not exist.

Why you cannot just trust the formatted output

Students sometimes ask an AI to "format these into APA" and assume that if the AI produces clean APA it must be pulling from something real. It is not. Formatting and factual accuracy are completely different problems for a language model. It can format a perfect citation for a paper that does not exist, and it can format a real paper incorrectly. The formatting tells you nothing about whether the paper is real.

For actual formatting, use a deterministic tool rather than a language model. Rubrica's citation formatter takes a real DOI or verified metadata and outputs correct APA, Harvard, MLA or Chicago without hallucinating anything, because it is not generating text, it is transforming structured data. Use the formatter for formatting. Use the verifier for existence.

The order of operations when you have used AI

If you have used ChatGPT, Claude, Gemini, or any other LLM at any point in your research workflow, run this sequence before submission. It takes about twenty minutes for a typical essay bibliography.

List every source that came from or was "confirmed by" an AI. Be honest with yourself. Include sources the AI only "summarised" for you.
Paste the full reference list into a verifier. The Rubrica reference verifier does this via CrossRef in one pass. Any source flagged as unresolved is a candidate for fabrication.
For each unresolved source, run the three manual checks. DOI, exact-title search, author publication list. Two failures out of three means the source does not exist.
Delete every fabricated source from your essay. Do not "guess a real paper" as a replacement, and do not try to rewrite the AI-generated claim around a different real paper. Remove the sentence, then decide whether you need to replace it with something real or whether the essay stands without it.
Reformat your remaining real sources. Run them through a deterministic citation formatter to ensure the APA / Harvard / MLA output matches what your department requires.
Re-read your essay with only the real citations in place. Often the argument held up because it was grounded in real reading. Sometimes the argument depended on a fake paper and collapses. Better to learn that at home than in the misconduct hearing.

What your university actually knows about AI citations

Tutors spot fake citations faster than students realise. Three reasons:

Markers are domain experts. If you cite a paper in their field that they have never heard of, they notice. The field is smaller than it looks.
Turnitin now checks reference validity. The AI detection feature is noisy, but the citation-verification features in the latest versions flag suspicious bibliographies.
Markers have started spot-checking DOIs. When AI fabrication became widespread in 2023-2024, academic integrity offices trained markers to resolve two or three DOIs per essay, picked randomly. One broken DOI is enough to trigger a full-bibliography audit.

The asymmetry is brutal. It takes you two minutes to verify a bibliography. It takes the misconduct panel two months to conclude that you did not. The tool that creates the problem is faster than the tool that detects it, but the panel's verdict is binding.

The difference between using AI to think and using AI to cite

Nothing in this guide is an argument against using AI at all. Using a model to help you brainstorm an argument, critique a paragraph, explain a concept, or restructure a messy section is fine at most universities, subject to your course's specific policy. The problem is narrow and specific: using AI as a citation source.

A practical distinction:

Good AI use - "Read this paragraph and tell me if the argument is clear." "Explain Gramsci's concept of hegemony in two sentences." "Rewrite this introduction to be punchier."
Dangerous AI use - "Give me three recent sources on x." "What does the literature say about y?" "Find me a paper that argues z."

The first category uses the model for what it is actually good at: working with the text you already have. The second category treats it as a search engine, which it is not. If you need sources, use a real database (Google Scholar, JSTOR, Scopus, your library's discovery service). If you need a source verified, use a real verifier.

Handling citations your supervisor gave you

Even citations from your reading list can occasionally be wrong. Your supervisor might have a typo in a year, a wrong page range, or a slightly misremembered title. This is rare but worth two minutes of checking, because a broken citation in your reference list looks the same to a marker regardless of whether you or your supervisor caused it. Verify your supervisor's list once at the start of research. Save yourself the re-chase later.

Worked example: verifying a suspicious reference list

A student shows me this partial bibliography for a media studies essay:

Boyd, d. (2014). It's Complicated: The Social Lives of Networked Teens. Yale University Press.
Tufekci, Z. (2020). Networked protest reconsidered: A decade later. New Media and Society, 22(8), 1321-1345.
Chen, L., & Park, H. (2022). Digital dissent and state repression: A comparative analysis. International Journal of Communication, 16, 3402-3425.
Nagle, A. (2017). Kill All Normies. Zero Books.

Run the three checks on each.

Boyd (2014). Book, real publisher, searchable on the Yale University Press website. Real.

Tufekci (2020). Google Scholar, exact-title search: zero results. Tufekci's publication list: no paper by that title in 2020. Her actual 2020 paper in New Media and Society exists but has a different title. Fabricated.

Chen and Park (2022). International Journal of Communication is open access, all issues are browsable. Volume 16 exists. The article does not. Fabricated.

Nagle (2017). Real book, real publisher, multiple reviews exist. Real.

Two out of four fabricated. The essay as drafted quotes both fake papers for substantive claims. Either the student removes those claims, or the essay needs new real sources to replace them. This is the diagnosis the reference verifier gives you in seconds, and it is the diagnosis that would otherwise arrive via a misconduct letter.

The self-audit you should do before every submission

Every source in your bibliography has either been held in your hand (book), read on a journal page (article), or resolved through a working DOI.
No citation entered the bibliography via a language model without being independently verified.
Every in-text citation appears in the reference list, and every reference list entry is cited in the text at least once.
Formatting is consistent and matches your department's required style.
Titles of cited works match, exactly, the titles on the actual publications.

This is a ten-minute audit. It is the single most important ten minutes of your pre-submission workflow, because fabricated citations are the one category of mistake that can end your academic year regardless of how good the rest of your essay is.

What to do if you realise you have already submitted fabricated sources

If you have submitted and only later discovered a citation is fake, the correct move is to email your tutor before they mark the essay, flag the citation, and explain what happened. This is painful but always better than being found out. Most universities treat a self-report as a mitigating factor. A tutor who finds the fake themselves has no reason to show that leniency. The difference in outcome between "the student emailed us first" and "we caught the student" is often the difference between a formal warning and a full misconduct hearing.

The broader habit

Verifying citations is one part of a larger pre-submission habit that used to be optional and is now essential. Reading the assignment brief carefully, decoding the rubric, checking citations exist, making sure the word count is in range, confirming references match in-text cites - this is the boring final lap of writing an essay and it is where marks quietly disappear. If you are running a rubric-aligned review of your draft as well, Rubrica's main grading tool will surface the same kind of structural issues that markers will see, before they see them. But citations are the single most urgent category to get right, because unlike a weak argument or a soft conclusion, fake citations do not just lose marks. They lose the whole essay.

Final thought

The confidence with which an AI will hand you a fake citation is the most dangerous thing about it. Nothing in the output looks wrong. The DOI is the right length. The authors are real. The journal is real. The title is the kind of sentence an academic would plausibly write. But the paper does not exist, and your marker will find out in under a minute with a DOI lookup. Verify before you cite, every single time. It is the cheapest insurance in academic writing.

If you are short on time and holding a twenty-source bibliography that has any AI involvement at all, run it through a CrossRef-backed verifier now. That single check is worth more than a week of revision, because a good essay with two fake citations is a zero. A mediocre essay with real citations is a mediocre essay. The floor matters more than the ceiling right now.