How to Spot Fake AI-Generated References in Student Work
By Tomas Lerpiniere · Updated 2026-10-21
Fabricated citations are one of the clearest integrity signals in AI-assisted coursework. A polished bibliography can still contain titles, authors, and DOIs that never existed. This guide gives UK teachers, tutors, and IQAs a fast, fair workflow for checking references without turning every essay into a tribunal.
In this guide
Feedback Flows tools in this guide
Most markers already know the feeling: a Year 12 Extended Project, a Level 3 vocational essay, or a university-style FE assignment cites five journals you have never heard of, all with neat DOIs and tidy page ranges. Sometimes the learner has discovered niche scholarship. Sometimes a chatbot filled gaps with invented sources. Distinguishing those cases is a referencing skill, not a guessing game about “AI style.”
Integrity work is stronger when references sit beside other evidence. Pair citation checks with the approach in evidence over AI detection scores and, where authorship is in doubt, with writing fingerprints and authorship checks.
1. Why generative AI invents references
Large language models predict plausible text. When asked for “recent peer-reviewed sources,” they often assemble author names, journal titles, years, and DOI-shaped strings that look right but do not resolve. The learner may paste the list in good faith, or may know the sources are unverified. Either way, the bibliography is not evidence until it is checked.
- Hallucinated works: Title and authors do not exist in Crossref, library catalogues, or publisher sites.
- Mangled real works: A real paper is cited with the wrong year, volume, or DOI.
- Stretched claims: The source exists, but the sentence in the essay says something the paper never argued.
2. Red flags that deserve a spot-check
| Signal | What to do |
|---|---|
| DOI returns 404 or resolves to a different article | Treat as high priority; ask the learner for the PDF or permalink they used |
| Journal name you cannot find after two search attempts | Check spelling variants; if still missing, flag as likely fabrication |
| Perfectly neat page ranges on every entry (e.g. 1–20, 21–40) | Not proof alone, but worth sampling against real issues |
| In-text citation with no bibliography match (or the reverse) | Often a process error; still fix before any integrity conversation |
| Americanised spellings and generic “studies show” phrasing around citations | Combine with writing history; do not sanction on style alone |
3. A five-minute verification workflow
- Sample, do not boil the ocean. Check three unusual entries first.
- Resolve identifiers. Paste DOIs into Crossref or the publisher resolver.
- Match metadata. Confirm title, authors, year, and venue.
- Open the source. Skim the abstract or relevant section.
- Record the outcome. Note what failed and what you discussed with the learner.
For a quick public pass on a pasted reference list, try the free citation checker. For class and institution workflows that also match in-text citations to the bibliography, use the citation and reference checker.
4. When the source is real but the claim is wrong
Fabrication is not the only failure mode. Learners (and chatbots) often cite a real paper to support a claim the paper does not make. That is a teaching opportunity: show how to quote carefully, use page numbers where required, and avoid “orphaned” citations that decorate a paragraph without carrying evidence.
Optional semantic alignment in institutional checkers can highlight possible mismatches. Treat those highlights as investigation prompts, not automatic misconduct findings—the same principle as evidence over raw detector scores.
5. Free checks and institutional tooling
- Free citation checker: Fast Crossref-oriented check when you only need to test a list.
- Citation and reference checker: Body-to-bibliography matching, style awareness, and exportable reports.
- AI writing detector: Quoted passage highlights when generation risk sits beside referencing risk.
- Writing fingerprints: Authorship consistency with the learner’s own baseline.
- Similarity scoring: Peer copying inside your organisation, which sometimes includes shared fabricated lists.
Check references before the conversation
Start with a free list check when you need speed, then move to the full reference checker when you need in-text matching and a record for moderation.
Try the free citation checker · Explore the reference checker
6. Having a fair conversation with the learner
Bring the broken entries, not a vibe. Ask how they found each source, whether they opened the PDF, and whether they can explain the claim in their own words. If they cannot retrieve the work, treat that as a process failure first. Escalate under centre policy when fabrication is deliberate or repeated.
Keep learner data handling aligned with your data commitments expectations, and document outcomes on the learner profile so the next marker is not starting from scratch.
Frequently asked questions
Conclusion
Fake references are checkable. Sample unusual entries, resolve identifiers, read enough of the source to test the claim, and record what you decided. Tools speed the first pass; teachers still own the integrity conversation.
Related guides
- Evidence Over AI Detection Scores Why a single AI percentage is a weak basis for a conversation, and what better integrity evidence looks like.
- Writing Fingerprints and Authorship Checks in UK Schools How UK schools, colleges and ITPs can investigate suspected AI work fairly with writing fingerprints, process evidence and JCQ-aligned workflow.
- Spotting Copying Between Learners How UK schools, MATs and training providers spot copying between learners using organisation-only similarity banks, fair workflows and clear policy.
Continue exploring