How Teachers Spot AI-Generated References in Coursework

Of all the signals that point to AI use in student coursework, the reference list is one of the most revealing - and one of the most overlooked. When a student submits an essay or coursework folder with a bibliography, teachers tend to skim it, check the formatting, and move on. But generative AI tools have a well-documented habit of inventing sources: plausible-sounding author names, real-looking journal titles, page ranges and years that fit the convention perfectly - and that do not exist. Learning to spot AI-generated references in coursework gives you a concrete, checkable piece of evidence that complements the more impressionistic "this doesn't sound like the student" instinct. This guide explains how fabricated citations arise, how to check them quickly, and how to combine that check with a likelihood score and a fair conversation.

Why AI Invents References

Large language models generate text by predicting plausible sequences of words, not by retrieving facts from a database. When a model is asked to support an argument with academic sources, it produces something that looks like a citation because it has seen millions of real citations and learned their shape. It assembles a believable author surname, a title that matches the topic, a journal that publishes in that field, and a year and page range that fit. The result reads convincingly. It is also, frequently, entirely fictional - what the research community calls a hallucinated reference.

This matters for coursework because students who use AI to draft an essay often paste the AI's references straight into their bibliography without checking them. They assume the tool retrieved real sources. A student doing their own research, by contrast, can only cite what they actually found - their references trace back to a library catalogue, a textbook, a website they can name. The gap between "references that exist" and "references that merely look right" is exactly the gap between genuine research and generated text.

Subjects with a formal referencing requirement - A-Level coursework with extended writing, EPQs, GCSE History and Geography controlled assessments, and any work asking for a bibliography - are where this tell is most useful. A piece with no references gives you nothing to check; a piece with eight neat Harvard-style citations gives you eight things to verify.

How to Check References Quickly

You do not need to verify every source in a class set. The aim is a fast triage that flags the work worth a closer look. Start with two or three of the most specific-looking citations - the ones with named journals, volume numbers, and page ranges. Search the exact title in a search engine and in Google Scholar. A real academic source surfaces almost immediately; a fabricated one returns nothing, or returns a real author who never wrote the cited piece, or a real journal that never published that title.

Watch for a few recurring patterns. Fabricated references are often too perfect: uniformly formatted, every field present, no missing page numbers or "n.d." entries that real student bibliographies usually contain. They sometimes pair a real, well-known author with a title that author never wrote, because the model associates the name with the topic. DOIs, when present, may not resolve. And the sources are frequently a little too on-the-nose for the argument, as if written to support the exact point the essay makes - because, in effect, they were.

A student's genuine bibliography tends to be messier and more human: a mix of a textbook, a couple of websites, perhaps one journal article they found and half-understood, with the odd formatting slip. That unevenness is a good sign, not a bad one. It reflects a real person doing real, imperfect research.

What the Likelihood Score Adds

Checking references tells you about the sources; it does not tell you about the prose. The two together are stronger than either alone. GradeOrbit's AI detection returns a score between 0 and 100 on the body of the writing - a measure of how far the piece departs from that pupil's own earlier writing on file. If the prose scores high and the references do not check out, you have two independent signals pointing the same way. If the prose reads as the student's own but one citation looks off, that is more likely a sloppy reference than wholesale AI use.

As we explain in our guide on how AI detection likelihood scores work, the number is one input into a picture, never a verdict on its own. The reference check is a second, very tangible input - and one a student can be asked about directly. GradeOrbit's detection, included with unlimited checks in every school teacher seat, works the same way whether that's lower-stakes work or coursework heading into moderation - there's no separate higher-confidence tier to reach for. It compares the writing against the student's own earlier work on file, which gives you a more grounded signal than a generic score alone.

Turning the Evidence Into a Fair Conversation

If both the prose and the references raise concerns, the next step is a conversation, not an accusation. References give you an especially good opening because they are concrete and specific. Ask the student to bring in or point you to one of the sources they cited. "Can you show me where you found this article?" is a simple, reasonable question that a student who genuinely used a source can answer, and that a student who pasted an AI-generated citation cannot.

Keep it low-key and curious rather than confrontational. A student who did the research will be able to describe the source, summarise what it argued, and explain why they used it. A student relying on fabricated citations will usually struggle to produce the source at all, or will admit they "found it online" without being able to say where. The reference check shifts the conversation from your subjective impression of the writing to a factual question about whether a specific source exists - which is far easier for both of you to discuss fairly. The broader framework for this sits in how teachers handle a high AI likelihood score fairly.

Detection should never replace your professional judgment, and a fabricated-looking reference is a prompt to investigate, not a conviction. Occasionally a student has simply mis-transcribed a real source or cited something from memory and got the details wrong. The conversation is what distinguishes an honest error from generated content - which is exactly why the conversation, not the tool, is where the decision is made.

Documenting the Check for Moderation

Whatever you conclude, record it at the time. For a reference-based concern this is straightforward and persuasive: note which citations you checked, what your searches returned, the likelihood score on the prose, what you asked the student, and what they said. "Three of six cited sources returned no results in Google Scholar or general search; prose likelihood score 81%; student unable to produce two of the sources when asked" is a clear, factual record that stands up well if the case is revisited.

For coursework heading into any form of moderation, include the detection run and your reference notes alongside your marking. A moderator reviewing borderline work benefits from seeing that you applied due diligence on both the writing and its sources - that evidence fits the wider pack for internally assessed components. If your school's academic integrity policy does not yet address AI specifically, the school AI academic integrity policy guide is a useful starting point.

Running Coursework Through GradeOrbit

AI detection in GradeOrbit is part of its school accounts and follows the same process for every subject. Run it from a pupil's profile for one piece, or on a whole class at once. Upload scans, photos or PDFs, or use the phone QR camera link; for a class set, bulk upload splits the pages by "pages per student", reads the name on each cover page and matches it to your class list, asking you to pick the pupil if the match isn't clear. Typed coursework goes in as a PDF or image (export Word documents to PDF first). The AI reads the pages and compares the work with that pupil's own earlier work already on file, returning a 0-100 score, a confidence label, the signals it noticed and a short reasoning paragraph on the pupil's profile. A pupil with no earlier work on file is skipped, and you are told.

School accounts run under a signed Data Processing Agreement: work is kept against the pupil, with their name, for the school's retention window, scans are removed when a run archives, and nothing is ever used to train AI models. Your reference check and your decision still belong in your own notes and your school's systems. For coursework where you also want to detect AI in the body of the writing subject by subject, our guide on detecting AI in A-Level English coursework applies the same principles to extended writing.

Try GradeOrbit for AI Detection in Coursework

If you want a reliable AI detection tool that gives you a clear likelihood score to work alongside your own reference checks and professional judgment, GradeOrbit is built for exactly that. Detection is included in every school teacher seat, so your department can check every piece that raises a question rather than rationing checks.

Start a free school trial to try detection on your department's real coursework.