How a citation gets invented

Watch it happen

A citation has a rigid form: authors, year, title, journal, volume, issue, pages, DOI. Rigid forms are exactly what a next-token predictor reproduces most easily. Every component is drawn from a strong pattern:

• Authors who publish in that area, in the right initials-and-surname format.
• A year in the plausible range for that literature.
• A title assembled from the vocabulary of the field, often a very good title.
• A journal that publishes that subject.
• A volume and page range consistent with that journal’s numbering.
• A DOI with a real prefix, since prefixes are publisher-specific and heavily represented, and a suffix that is invented.

Every part is individually plausible. The combination has never existed. This is not a bug in the citation feature: there is no citation feature. It is the same process that writes the prose, applied to a highly structured string.

The hybrid, which is worse

Purely invented references are the easy case; a search finds nothing and you delete it. The dangerous ones are hybrids:

• A real paper with a wrong year, which breaks your reference list and looks like carelessness.
• Real authors, real journal, title that does not exist.
• A real paper attached to a claim it does not make. This is the worst, because every element checks out and the citation is still false. Only reading the paper catches it.
• A real paper whose findings are reversed: the study found no effect, and the citation is attached to a sentence saying it found one.

The last two are why the reference exists is not the same as the reference supports this. Two separate checks.

Why asking the model does not help

Is this a real paper? is a question about the world, and the model has no channel to the world. It answers from the same process that produced the citation. It will often say yes. Occasionally it will say no about a real paper, which is even more confusing.

This holds when search is switched on, in a modified form. Now it may retrieve something, and the failure becomes a real link attached to a claim the page does not make. Which brings us to the only rule that survives every version of this technology.

The rule

Open it. If you did not open it, you did not check it.

Open means: reach the actual paper or book, find the page, and read the sentence that supports your claim. Not the abstract if your claim is about a method. Not a search result snippet. The thing itself.

This takes about three minutes per source with the free tools in the next lesson. For a ten-source essay that is half an hour, and it is the half hour that stands between you and the most common academic-misconduct finding in the world right now.

The DOI test, which takes ten seconds

Every real journal article since roughly 2000 has a DOI. Paste it after https://doi.org/ and press enter. A real DOI resolves to the publisher’s page for that article. An invented one gives you a DOI-not-found error.

This single check disposes of most fabricated references instantly. It does not tell you whether the paper supports your claim, only that it exists. Both checks are required and this is the cheap one.

What a supervisor sees

The reason this matters more than it feels like it should: an academic reading your bibliography recognises the literature. They know the journals, often the authors, sometimes the specific paper. A fabricated citation in a field somebody has worked in for twenty years is as visible as a misspelled name in your own language.

And the inference they draw is not this student used AI. It is this student cited something they never read, which has always been misconduct, long before any of this existed. That is the charge, and it is a fair one.

Using a model for references honestly

There is a legitimate version, and it is a search-strategy conversation rather than a citation request:

What are the main research areas relevant to this question, what search terms would find them, and which journals publish this kind of work? Do not give me citations.

Then search Google Scholar, PubMed or arXiv with those terms yourself. The model is good at knowing the shape of a field and at translating your vague question into the vocabulary a database will match. It is not a bibliography. Used that way, it saves genuine time and generates nothing you then have to un-invent.