Why AI Invents Sources in the First Place
A language model does not look up references. It predicts what a plausible citation looks like — a believable author, a journal that publishes in your field, a year that fits, a DOI-shaped string of characters. The result reads perfectly and points to nothing. This is the single fastest way to torch your credibility with a professor, because checking one citation takes them thirty seconds and checking yours takes you zero.
Three failure modes show up constantly:
- Fully fabricated — the paper does not exist. Confident title, real-sounding authors, dead end.
- Frankenstein — a real author stapled to a real journal stapled to a title none of them wrote.
- Misattributed — the source is real, but it does not say what the AI claims it says. The most dangerous kind, because the link works.
The rule that saves you: the model can suggest where to look, never what to cite. A suggestion is a lead. A citation is a thing you have personally opened and read.
The Verify-Every-Source Workflow
Treat every AI-surfaced reference as guilty until proven real. Run each one through four checks, in order, and stop the moment it fails:
- Does it exist? Copy the exact title into Google Scholar or your library database — in quotes. No quoted-title match means it is fabricated. Delete it.
- Do the details match? Confirm authors, year, journal, and volume against the actual record. A real paper with a wrong year is still a wrong citation.
- Does it say what you claim? Open the source. Find the sentence or finding you are leaning on. If you cannot point to it, you cannot cite it.
- Is the DOI live? Paste it after
https://doi.org/. If it 404s, the identifier is invented even when the paper is real.
Make the AI do the legwork for verification, not for trust:
Here is a reference you suggested:
[paste the full citation]
Do not confirm it from memory. List the exact search query
I should run in Google Scholar to locate this source, and tell
me which specific details (authors, year, journal, volume, DOI)
I need to match against the real record.
A tool with live retrieval — Consult-style assistants, Elicit, Scite, or a search-grounded model — narrows the fabrication rate but does not zero it. Grounded tools still misattribute claims and still hand you a real DOI attached to the wrong sentence. Steps 3 and 4 are non-negotiable regardless of how fancy the tool is. The full source-vetting routine is drilled in AI for Academic Research & Papers if you want the extended version.
Don't Let AI Format Your Citations Blind
AI is genuinely useful for the mechanical part of citations — converting a verified reference into APA, MLA, Chicago, or your department's house style. The catch: it will format with the same confidence whether the underlying source is real or invented. So formatting comes last, only after a source clears all four checks.
Feed it clean inputs and it does clean work:
Format this source in APA 7th edition. I have already verified
every detail against the published record — do not change any
author name, year, title, or page number. Only fix the format:
Author(s): [...]
Year: [...]
Title: [...]
Journal, volume(issue), pages: [...]
DOI: [...]
Even then, eyeball the output. Models routinely drop the second author, italicize the wrong element, or "helpfully" correct a year that was right. Cross-check the formatted result against your style guide's example for that source type — book chapter, preprint, and dataset citations are where it slips most.
For the original record itself, prefer the publisher's own "cite this" export or a managed library like Zotero. Pull the BibTeX or RIS straight from the source, then let AI translate the style if you need a format your manager doesn't support. That keeps the facts coming from the publisher and uses AI only for the typography.
Build a Paper Trail You Can Defend
The cheapest insurance against an integrity meeting is being able to show your work. As you verify, keep a running log — a spreadsheet column, a Zotero note, anything — with: the claim, the verified source, the page or section it came from, and the date you confirmed it. When a professor asks "where does this number come from?" you answer in ten seconds instead of re-deriving it under pressure.
Two habits make this bulletproof:
- Cite from the source, not the summary. If AI summarized a paper for you, the citation still has to point to the paper you opened — not to the chat. The summary is scaffolding; the source is the evidence.
- Quarantine the unverified. Keep AI-suggested leads in a separate list from confirmed citations. Nothing migrates into your bibliography until it has passed all four checks. A reference that is "probably fine" is a reference you have not checked.
This discipline is also where citation accuracy stops being a formatting issue and becomes an honesty issue — the line between using a tool and misrepresenting your sources. If that distinction feels fuzzy, AI Ethics & Responsible AI maps it out cleanly.
The One-Sentence Standard
Hold every citation in your paper to this test: you have opened the source, found the exact claim it supports, and confirmed its details against the published record. If a reference cannot pass that sentence, it does not go in — no matter how good it looks or how sure the AI sounded.
Do this and AI becomes what it should be: a faster way to find real sources and format them correctly, never a generator of plausible fiction with your name on the byline.