Sources and citations
~ min read
30-second summary
- AI invents citations that “sound real”: plausible author, plausible journal, none of it actual. Four typical fabrication patterns to spot at a glance.
- The DOI is the fastest filter: ten seconds per citation. If it doesn’t resolve on
doi.org, the citation is almost always made up. - Three filters to tell a real academic source from a dressed-up blog: resolvable DOI, journal indexed in Scopus or Web of Science, known publisher.
- Four main styles: APA, MLA, Chicago, ISO 690. AI formats them well if you give it the raw data, but it’s not a reliable source for finding them.
- For a closed bibliography (papers you’ve already selected), use NotebookLM or Claude Projects: they only answer from the documents you’ve uploaded, no invented citations by design.
The academic bibliography is where AI does the most damage. Not because it gets things wrong more often than elsewhere, but because it has learned to “sound like a bibliography” very well without necessarily producing a real one. The citations it offers have the right shape: an author with a plausible surname, a journal that exists, a recent year, credible page numbers. The specific combination just often doesn’t exist.
Telling a fake citation from a real one is a skill, and it can be learned in one sitting. The practical consequence of not learning it: in thirty seconds your supervisor picks a random citation from your bibliography, looks up the DOI, and figures out that you’ve handed in citations that never existed. No drama, it’s predictable and preventable. This lesson is the prevention.
Anatomy of a fabricated citation
Section titled “Anatomy of a fabricated citation”AI doesn’t invent with “tags” that let you unmask it. It’s been trained on millions of real citations and has learned the formal pattern very well, but “producing a plausible citation” and “producing a citation that exists” are the same task as far as it’s concerned: which combination actually exists is information external to the way it writes, sitting in a database it doesn’t query while it responds. So it invents with the same formal structure as real citations: an author with a surname that really exists, sometimes a well-known academic in the field; an existing journal or close to it (sometimes with a small distortion of the name); a plausible year, usually recent; credible page numbers. What’s missing is the match: the specific author + title + journal + year combination doesn’t exist, or the DOI doesn’t resolve, or the article exists but says something different from what the citation attributes to it.
Four typical fabrication patterns worth keeping in mind.
Real author + real paper + wrong year. The author exists, the paper exists, but the citation dates it to 2018 when it’s actually from 2014. Mean variant: the year is when you thought you read something similar, and you believe it.
Real author + non-existent paper attributed to that author. A well-respected scholar in the field, cited as the author of an article they never wrote. The title is plausible (“Tax Pass-Through on Sugar-Sweetened Beverages in European Markets”), and you believe it because you know the author.
Real journal + non-existent issue/year. The Journal of Health Economics exists, but volume 67 issue 3 from 2019 doesn’t contain the article AI is citing. The journal is real, the article isn’t.
Perfectly structured but one hundred percent fabricated citation. Invented author, invented paper, invented journal (or one that happens to exist by chance), invented DOI. The form is impeccable. This is the most frequent case when you ask AI for “three recent papers on X” without access to a live search engine.
The DOI as a verification gate
Section titled “The DOI as a verification gate”Every serious academic article published since roughly 2000 has a
DOI, Digital Object Identifier. It’s a unique signature for the
article, a string that starts with 10. followed by a publisher
identifier, a slash, and a code. Example:
10.1016/j.jhealeco.2019.102225 is the DOI for an article by Cawley
et al. from 2019 in Journal of Health Economics, which you’ll see
in the ChatDemo further down.
The practical check is fast. Take the DOI AI gave you, paste it onto
doi.org/<DOI> (like this: https://doi.org/10.1086/675329), or
look it up directly on Google Scholar. Three possible outcomes.
The DOI doesn’t resolve. The page says the document doesn’t
exist, or doi.org sends you to an error. The citation is almost
certainly made up. Throw it out.
The DOI resolves, but to a different paper than expected. Open the link and read title, authors, year. If they don’t match what AI cited, AI has attributed the wrong content to a real identifier. Throw it out.
The DOI resolves to the right paper. Everything matches: authors, year, title, journal. The citation is almost certainly real. What remains is checking that the paper actually says what AI attributes to it as content: for that you need at least the abstract.
The DOI is the fastest filter you have: ten seconds per citation. For pre-2000 articles or books (which often don’t have a DOI), verification goes through WorldCat or your library’s catalog: you search author and title, and if nothing comes up the citation needs checking more carefully.
Four citation styles
Section titled “Four citation styles”Four main styles circulate in academic work, depending on the field. Short map.
APA (American Psychological Association). Social sciences and
humanities: psychology, sociology, economics, education. In-text
citation (Author, year), final bibliography ordered by author’s
surname.
MLA (Modern Language Association). Anglo humanities: literature,
languages, cultural studies. In-text citation (Author, page), no
year. The bibliography is called “Works Cited”.
Chicago (notes + bibliography). History, some Italian humanities. Footnotes with full citation the first time, shortened afterwards. Separate bibliography at the end.
ISO 690 (with its European variants). Technical sciences and
theses in Italian, often in engineering and some science
departments. Numbered references [1], [2], ... in the text,
bibliography at the end in the same order.
If your supervisor doesn’t specify, ask: one email line settles it. With no guidance, follow your department’s convention. Economics, psychology, sociology usually APA. English literatures and language studies MLA. History and Italian humanities Chicago. Engineering and technical sciences ISO 690. When in doubt, leaf through two or three recent theses your supervisor has supervised (ask the academic office) and use the style they used.
For any of these four, AI is a good formatter: you give it the raw data (author, title, year, pages, journal) and ask “format these five entries in Chicago style”. The result is usually acceptable. What it isn’t good at is remembering the exact rules of each style: every version gets updates, and AI can get a detail wrong (comma instead of period, italics on the journal title instead of the book title). To check whether the formatting is right, use an official guide (Purdue OWL is the strongest free reference in English) or, better, a bibliography manager like Zotero (free, open source). In practice: you install the desktop app, add the browser extension, and when you’re on a paper’s page you click an icon that imports author, title, year, journal, DOI in one move. From there you format in any style (APA, MLA, Chicago, ISO 690) without asking AI for anything. Free cloud sync with an account, plugins for Word and Google Docs.
Real academic source or dressed-up blog
Section titled “Real academic source or dressed-up blog”When you ask AI to “find me sources on X”, the risk isn’t only fabrication. It’s also a confusion of register: it offers you things that look academic but aren’t, as if they were academic sources.
Three typical cases:
- articles on
medium.comwith a scholarly tone and internal citations; - blogs by doctoral students or research centers, well written but not peer-reviewed;
- Draft of an academic article posted online before peer review. It can be revised or withdrawn, but it saves the months of waiting for journal publication. on SSRN or ResearchGate not yet published in a journal, which can still be modified or withdrawn.
None of this is “junk”: a good blog by a doctoral student can be informed, a preprint can be a serious piece of work waiting for publication. But using them in a thesis as if they were peer-reviewed is a mistake of register that your supervisor will flag immediately.
Three practical filters to tell them apart.
-
Resolvable DOI. The gate seen above. If there’s no DOI, or there is one but it doesn’t resolve, the source doesn’t pass the academic filter.
-
Indexed journal. Search the journal title on
scopus.comor Web of Science (free through your library’s proxy: the system that lets you, from the campus Wi-Fi or from home logging in with student credentials, access databases your university subscribes to). If the journal is Included in an official academic database (Scopus, Web of Science) that certifies a journal as peer-reviewed and recognized in its field. Not every 'serious' journal is indexed, but indexed journals are almost always serious. in at least one of the two, it’s a peer-reviewed journal recognized in its field. -
Known publisher. Springer, Elsevier, Cambridge University Press, Oxford University Press, JSTOR, Taylor & Francis, the university presses of major universities. If the publisher is an unknown and the DOI doesn’t resolve, be suspicious.
Two filters out of three passed: the source is probably solid. Zero filters out of three: you’re citing a blog, not a paper.
Three papers on taxation and food consumption
Section titled “Three papers on taxation and food consumption”Economics student, working on a term paper. Asks AI for three recent papers on a specific topic. AI replies with three well-formatted citations, APA style, complete with DOIs.
The student opens doi.org and tries the three DOIs. The first
doesn’t resolve: error page, article doesn’t exist. The second
resolves, but opens a paper on a completely different topic,
something about electric cars and public incentives in Norway. The
third resolves correctly and opens exactly the Cawley et al. 2019
article on the Philadelphia beverage tax, which really does talk
about taxation on food consumption. Two of the three citations are
made up: one with a fake DOI, one with a DOI that leads somewhere
else.
The next move, to paste into the same chat: “I’ve verified. The first DOI doesn’t resolve, the second opens a paper on a completely different topic. Let’s only work on the third, Cawley et al. 2019. Add me three recent papers that cite or extend that one, and this time give me verifiable DOIs or tell me if you don’t have any with certainty.” From here the conversation realigns: AI has a real anchor point (a verified paper) and an explicit safety exit (it can say “I don’t have reliable recent sources”). It’s the same pattern already seen in Research for work: give AI an exit to admit the limit, and it invents less.
What NOT to do
Section titled “What NOT to do”Don’t hand your bibliography to the supervisor without
verifying every DOI. In thirty seconds they pick one at random,
search, and unmask you. Opening ten links on doi.org costs you
two minutes, and it’s the difference between a thesis you can hand
in and an embarrassing moment in the middle of drafting.
Don’t use “find me recent papers” as a primary source. For real research you start from Google Scholar or your university’s database (Scopus, JSTOR, ProQuest, MLA International Bibliography depending on the field), not from AI. AI is fine as a second step, when you already have a core of papers and you want to format, synthesize, or look for connections.
Don’t trust AI when it says “this paper argues X”. Even if the DOI resolves to the right paper, AI can attribute a claim to the paper that the paper doesn’t support. Verify by reading at least the abstract; for papers that will actually go into the bibliography, read the paper.
Check what you’ve understood
Section titled “Check what you’ve understood”What comes next
Section titled “What comes next”Bibliography handled. The next lesson is a side topic but useful if you do languages or classics: translating and comparing texts.