Research, And Sources That Exist

The failure that keeps happening

In May 2023, a New York lawyer named Steven Schwartz filed a court brief containing six cited cases. Names, docket numbers, judges, quoted passages. All six were invented by ChatGPT.

The part worth memorising is what he did when opposing counsel could not find them. He went back to the chatbot and asked whether the cases were real. It said yes. He asked for confirmation. It produced full text. He filed that too. The judge fined him and his colleague $5,000.

Students do the smaller version of this every term, and it is far easier to catch than AI-written prose, because a marker who wants to check a reference simply checks it. A fabricated bibliography is one of the few genuinely reliable signals a tired academic has.

Why citations in particular

A language model, absent a search tool, is producing statistically plausible text. A citation is an unusually learnable format: author surname, year, a title made of field-appropriate nouns, a journal that publishes in that field, a volume number, a page range, a DOI-shaped string. Every component is drawn from a pattern the model has seen thousands of times.

The result is a reference that is right in every respect except existing.

This is not the model lying. There is no database being consulted and no lookup being fumbled. Unless the tool is actually searching — and you should always know whether yours is — nothing in the process ever checks reality.

The rule, and how to satisfy it

Never cite a source you have not opened. Not summarised, not confirmed. Opened.

How to check, in order, ninety seconds each:

1. 1Search the exact title in Google Scholar, your library catalogue, or your discipline's index — PubMed, JSTOR, SSRN, arXiv. If the exact title returns nothing anywhere, stop. It does not exist.
2. 2Resolve the DOI. Paste it after https://doi.org/. A real DOI lands on the paper. An invented one returns an error. This takes five seconds and catches most fabrications.
3. 3Check author, year, journal and volume together. Fabrications frequently attach real, well-known researchers to a paper they never wrote. Finding that Bjork is a real memory researcher tells you nothing about whether this particular Bjork paper exists.
4. 4Open it and find your sentence. This catches the more common failure, which is not invention at all: a real paper cited for a claim it does not make, or makes about a different population, or explicitly argues against.

That last one is worth dwelling on. Genuine papers get miscited more often than fake ones get invented, and miscitation is the error that survives all the way into your finished work.

Search-enabled tools change the risk, not the rule

More tools now retrieve and link — ChatGPT with browsing, Perplexity, Gemini, Claude with search, Copilot. A link you can click is genuinely different from a fabricated reference, and this is real progress.

But the summary can still drift from the source. The tool retrieves a page, compresses it, and the compression may sharpen a hedged finding into a firm one, or lose the sample size, or attribute a result to the wrong study among several on the page. Click through. Read the paragraph. It is quick, and it is the difference between citing and gesturing.

Also notice what the retrieved sources actually are. A confident answer built on three blog posts and a press release is not the same as one built on the paper.

Where AI genuinely helps with research

Use it for the map, not the territory.

What are the main competing positions on whether minimum
wage increases reduce employment? Who argues each side,
and what is the strongest objection to each?

I'm researching informal waste collection in Indian cities.
What search terms would a subject librarian use that I
probably wouldn't think of?

Here's the methods section of a paper. Explain the
regression discontinuity design in plain terms so I can
follow the results.

All three are things it is good at, and none produce a citation. You get orientation, vocabulary, and comprehension — then you go and read.

One more, which will save you an afternoon: after you have your source list, ask it what you appear to have missed, or which viewpoint is absent from your reading. A bibliography that only contains people who agree with you is a common, avoidable mark deduction.