A bibliography can look reassuringly scholarly even when one of its entries has never existed. Picture a student preparing a geography project. An AI assistant proposes a journal article with two authors, an impressive title and an exact page range. The reference looks complete enough to paste into the final report. But the student has not found the article itself. What should happen before that citation becomes evidence?
Citation verification means checking that a referenced work really exists, that its authors, title, publication details and identifier match a trustworthy record, and that the source actually supports the statement for which it is cited. The immediate answer is: treat an AI-suggested reference as an unverified search lead, find an original publisher or catalogue record, inspect the work where possible, and cite only what you can trace. Fluent formatting is not verification.
Did you know? A real-looking reference has several separate claims
A reference is not merely a string of punctuation. Its title asserts that a work with that name exists. Its author field asserts responsibility. Its year identifies a publication version. A DOI or other stable identifier asserts a particular record. The student’s accompanying sentence makes a further claim: that the identified work contains relevant evidence. One of these can be correct while another is wrong.
For scholarly works that have Crossref DOIs, Crossref’s own metadata documentation describes how deposited identifiers and bibliographic fields can be searched. A match is useful evidence of a registered item, but the record alone does not prove a student has read the paper or that its findings support their paragraph. Not every legitimate publication has a Crossref DOI; absence from one index is a reason to investigate further, not proof that a work is invented.
A worked example: the bibliography entry that must stay in quarantine
Invented classroom example: imagine an AI tool supplies the following alleged reference: “Tan, R. and Vale, P. (2025). Universal Recall Gains From Colour-Coded Homework. Journal of Student Study Methods, 18(2), 41–58.” The author names, journal, title and numbers in this example are fictional. They are not a real source or a claim about research results. No DOI is provided. The student wants to write: “Colour-coded homework improves memory for every learner.”
- First test — existence: search the exact alleged title in quotation marks, then try its distinctive keywords in a relevant library catalogue, scholarly index or publisher website. Record whether the alleged journal issue and article can actually be located.
- Second test — identity: if a similar-looking publication appears, compare title, authors, year, journal, volume, issue, pages and identifier one field at a time. A neighbouring paper is not a match merely because its topic sounds similar.
- Third test — access: if a genuine work is found, open its abstract or text through a legitimate route and identify precisely what the researchers examined. A headline or catalogue entry is not a substitute for the evidence.
- Fourth test — support: the word “every” demands evidence about every learner, not simply an average improvement in one group. If the actual work supports only a limited association or a selected sample, narrow the sentence accordingly.
- Fifth test — honest outcome: if no reliable record can be located, remove the alleged reference and unsupported claim from the bibliography. Mark the source lead unverified in private research notes rather than inventing missing fields.
This is a worked decision procedure, not a claim that a specific invented article failed a real search. The strongest possible outcome is not always finding something to cite. Sometimes it is preventing a fictitious reference from entering the report.
The small diagnostic that locates the student’s actual difficulty
Give a learner three source cards: Card A includes an authentic-looking reference but no accessible original; Card B points to a verifiable publisher page whose author and date match; Card C points to a genuine article but the proposed student sentence exaggerates its findings. Ask which cards are safe to cite for that exact sentence. A careful learner says A is unverified; B still needs a relevance check; and C requires a narrower claim. This distinguishes three skills: finding a source, matching its identity and evaluating what it proves.
Watch for different misconceptions. A learner who believes any DOI-shaped text is valid needs identifier checking. One who finds a paper with a similar title but stops there needs field matching. One who verifies every bibliographic detail yet overstates the study’s claim needs reading and inference practice. Repeating citation punctuation lessons would miss all three needs.
Practice: supported search, independent check, then transfer
Round one: a modelled source trail
The teacher demonstrates a verified record by starting on a publisher page, copying the title and author names exactly, opening the abstract, and identifying one sentence the source can support. Then the class compares the copied record with the original. If a DOI is supplied, follow its resolver to the expected publication rather than accepting the printed DOI on sight.
Round two: inspect a deliberate mismatch together
Prepare two fictional practice cards: one says a paper was published in 2024 with two named authors, while the accompanying mock publisher record says 2022 with three authors. Students should not choose whichever year looks neater. They must point out the mismatches, refuse to merge two records, and seek the correct source identity before formatting a citation.
Round three: verify a new reference unaided
Provide a genuine, accessible article chosen by the teacher and a deliberately inaccurate citation for it. Ask the student to correct only fields verified against the source, then explain what one paragraph of the article actually establishes. Success requires both an accurate reference and a supported sentence. Do this with approved materials rather than sending children to random or paywalled sources.
A five-column research log that makes uncertainty visible
- Proposed citation or source lead
- Original record actually opened (publisher, library catalogue, archive or stable repository)
- Fields confirmed, with a note of any mismatches
- Exact claim or passage checked and its limitation
- Status: verified for this claim, verified item but claim unverified, or not located
Keep the proposed reference separate from the verified version; do not overwrite the trail of what was checked. For an assignment requiring a particular style, apply that style after verifying factual fields. Formatting rules vary between schools and manuals, but the existence of a source is not a style preference.
Where AI, libraries and human judgement each belong
AI can propose search terms, suggest possible authors to investigate, or help a learner generate a source-checking checklist if the teacher permits that use. A catalogue or publisher provides evidence about a publication’s identity. The source itself provides material for evaluating the claim. The learner remains responsible for deciding whether and how to use it. The University of Washington Law Library’s verification guidance also emphasises checking AI-generated citations against actual sources; its legal research examples are specialist, but the verify-before-relying principle transfers to school research.
A parent can ask, “Can you open the work you listed, and point to what you used?” A teacher can permit an honest “not located” mark and provide accessible source packets so the task evaluates research judgement rather than access to expensive databases. Students should follow their school’s rules for declaring AI assistance and should avoid uploading private school documents into unapproved tools.
What progress looks like—and what it does not
A practical progress record notes how many candidate references the student independently matched to originals, how many mismatches they detected, and whether each cited sentence is actually supported. For example, in an invented exercise with five supplied candidates, identifying four correctly yields 4 ÷ 5 = 80% on that particular diagnostic. It is not evidence of 80% accuracy on real research, nor a guarantee of future marks. Follow up with a fresh source in a different subject.
Frequently asked questions
Does a working DOI guarantee a paper is trustworthy? No. It helps identify a registered object. Research quality, corrections, retractions, relevance and the exact claim still require checking.
What if the source is behind a paywall? Look for lawful library access, a publisher abstract, an institutional repository copy or an alternative openly available source. Do not pretend to have read text you cannot access.
What if the AI gives a real author but the wrong article title? Treat the full reference as unverified. Do not silently combine details from different publications to manufacture a record.
Does every school report need journal articles? No. Official documents, books, primary records and other suitable materials may be stronger for the question. Choose sources for relevance and quality rather than impressive journal formatting.
A better bibliography is one a reader can retrace
A learner who verifies the document, its identity and the claim can explain their evidence without the AI tool present. For adjacent lessons, see how to decide whether a source is primary for a history question and how to separate evidence from inference. Continue through the eduKate Education Hub. The unchanged Clementi Mathematics learning reference illustrates the same sequence: diagnose the gap, explain, practise, and verify independently. Real education is not a longer bibliography. It is a reference trail that can survive a reader’s next question.