Practical guidance on citing AI use, verifying AI-generated references, and academic honesty.
Module · Research + Writing
Lesson 9 of 9 available lessons
⬡ What you'll build
A fabricated citation is the most damaging thing AI can slip into your work. It looks authoritative, it passes a glance, and it can turn an honest assignment into an integrity case — because you've cited a source that does not exist or does not say what you claimed. This lesson is the one place in the track that owns citation verification: a concrete, repeatable process for making every reference in your work real, correct, and honestly used.
If you haven't yet, read Spotting Hallucinations & Verifying AI Output first — this lesson goes deeper on the citation-specific half of that problem.
Citations are claims: "this specific source supports this specific point." When a citation is fake or misapplied, three things happen at once:
The standard is simple and non-negotiable: every citation in your work resolves to a real source you have opened, and the source actually says what you claim it says.
A language model predicts plausible text. A citation has a very predictable shape — author, year, title, journal, volume, DOI — so when the model doesn't have a real reference, it generates one that fits the pattern perfectly. The result reads exactly like a genuine citation because it was built from the statistical shape of thousands of real ones.
This is why fabricated references are more convincing than other AI errors, not less: the format is flawless even when the content is invented. You cannot detect them by looking — only by checking.
The fabrications come in five recognisable forms. Learn them so you know what you're checking for:
| Failure type | What it looks like | How you catch it |
|---|---|---|
| Fake author | A plausible name who never wrote the paper | Author + title don't co-occur in any database |
| Fake journal | An official-sounding journal that doesn't exist | No publisher, no ISSN, no journal homepage |
| Wrong year | Real paper, but the year/volume is off | Metadata doesn't match the real record |
| Invented DOI | A correctly-formatted DOI that resolves to nothing | doi.org returns an error |
| Real paper, wrong claim | A genuine source attached to a point it doesn't make | You open it and the claim isn't there |
The last one is the sneakiest — the citation is real, so a DOI check passes, but it's still wrong because it doesn't support your point. That's why verification has two halves: does it exist and does it support the claim.
Run every AI-suggested reference through this sequence. Stop and discard the moment a step fails.
Search the exact title
Resolve the DOI
Match the metadata
Confirm the journal is real
Open it and find the claim
Cite from the real record
⚠The two-part gate
A citation must pass both tests: (1) it exists (steps 1–4) and (2) it supports your claim (step 5). A real DOI is necessary, not sufficient.
If any box is unchecked, the reference does not go in your work.
Verification also means citing the right source. AI often hands you a secondary source (a review, a news article, a textbook) summarising a finding — but academic work usually wants the primary source where the finding was first reported.
| Primary source | Secondary source | |
|---|---|---|
| What it is | Where the research/data was first reported | A summary or commentary on primary work |
| Examples | Original study, dataset, court ruling, the actual text | Review article, textbook, news write-up, encyclopedia |
| Cite it when | Making a specific evidential claim | Giving background or an overview |
| AI risk | AI may invent it | AI may cite the summary as if it were the finding |
Rule: if you're claiming a specific result, trace it to and cite the primary source — don't cite the review that mentioned it. (AI is useful for finding the primary source via the secondary one — then you verify and cite the primary.)
See Writing with AI as a Partner for the disclosure norms around AI-assisted writing (this lesson covers the citation side).
If your assignment allows AI use and asks you to cite it, follow your required style guide — the major ones (APA, MLA, Chicago) have published AI-citation guidance, and institutions often have their own variant. General principles that hold across styles:
I used AI on this assignment in these ways: [list exactly what you did — e.g., brainstormed search terms, critiqued my draft, suggested counter-arguments]. I independently verified all sources. Draft a 2–3 sentence AI-use disclosure statement in plain, honest language that I can adapt. Do not overstate or understate what I did.
Always confirm the exact format against your institution's policy and required style guide — citation rules for AI are still evolving and differ between courses.
Example A — the invented DOI. A reference looks perfect: real-sounding authors, a known journal, a formatted DOI. The exact-title search returns nothing, and the DOI fails to resolve on doi.org. Both checks fail → the reference is fabricated and is discarded. Total time: under a minute.
Example B — real paper, wrong claim. A student cites a genuine, DOI-resolving paper to support "X causes Y." Opening it, the paper only reports a correlation and explicitly says it can't establish causation. The citation exists but doesn't support the claim → the student softens the claim to match what the source actually says.
Example C — secondary cited as primary. AI suggests a textbook for a specific statistic. The textbook is real, but it's quoting an original study. The student traces the statistic to that original study, verifies it, and cites the primary source instead.
AI fabricates references because a citation has a predictable shape it can fill convincingly. You cannot spot fakes by reading — only by checking. Run every reference through the two-part gate: does it exist (title + DOI + metadata + journal) and does it support the claim (open and read it). Cite the primary source for specific findings, disclose your AI use honestly per your institution's rules, and never let the AI's formatted text be your final citation. Verified sourcing is what makes your work both honest and defensible.
Exercise 1 — Verify 5 AI-generated citations. Ask an AI for 5 references on a topic in your field, then run each through the verification workflow. Success criteria: for all 5 you can state "verified — exists and supports the point" or "discarded — failed at step N," with the doi.org / Scholar result that proves it.
Exercise 2 — Find the fabricated reference. Get a list of 6–8 AI references and assume at least one is fake. Identify the fabricated one(s). Success criteria: you correctly flag the fabricated reference and show the failed check (no exact-title match or non-resolving DOI) — not a guess.
Exercise 3 — Does the source support the claim? Take one AI claim with its cited (real, DOI-resolving) source. Open the source and check. Success criteria: you can quote the passage that supports the claim — or show the source doesn't support it and rewrite the claim to match what it actually says.
Spotting Hallucinations & Verifying AI Output
The broader verification skill — facts, quotes and numbers, not just citations.
AI-Assisted Research Workflow for Students
Where citation verification fits the full discovery-to-synthesis pipeline.
Writing with AI as a Partner, Not a Ghost
The disclosure and authorship norms for AI-assisted writing.