AI that writes research with citations you can check

A language model writing a literature review produces citations the way it produces everything else: by generating text that fits the pattern. A plausible author, a plausible year, a plausible journal. The result reads exactly like a real reference and frequently is not one — the paper does not exist, or it exists and says something else.

The damage is worse than a wrong fact, because a citation is a promise the reader can check. Most will not. The one who does finds a source that does not support the claim, and from that point nothing else in the document is trusted either.

The question is not whether a tool produces citations. It is whether every number in the text can be traced to a sentence that appears verbatim in a paper you can open.

• Where did it search? Named databases — OpenAlex, Crossref, Semantic Scholar, Europe PMC — or the open web? A tool that will not say has not really searched.
• How many did it find, and how many did it use? A reproducible review reports both. “80 studies identified, 27 cited” is a claim you can audit. Silence about the first number usually means the second is all there was.
• Is every figure attributed to a specific paper? Not to the bibliography in general — to one entry. A headline number with no reference beside it is the commonest failure in AI-written research, and the hardest to spot because the prose reads well.
• Does any in-text marker point at nothing? A model can emit [47] in a paper with 30 references. Ask whether the tool checks.
• Does it review its own output? Checking each claim against the evidence is a different question from checking whether the paper contradicts itself between sections — and a paper can pass the first sentence by sentence while concluding the opposite of its own results.

The strongest check is mechanical rather than clever. For every number the tool extracts, it should also capture the exact sentence it was read from, then confirm two things before that number is allowed anywhere near the text: the sentence appears in the source material, and the number appears in the sentence.

This is deterministic. It does not ask a model whether it was honest — it checks. Anything that fails is dropped rather than flagged, because a figure that cannot be traced should never reach a table or a chart, and downstream review demonstrably does not catch it.

It also solves attribution for free: the match that proves a number is real is the same match that says which paper it came from, so the two cannot disagree.

A tool searching the open web will cite blog posts, press releases and preprints indiscriminately. For coursework or a thesis that is usually disqualifying. Searching scholarly databases means the reference list is made of things a supervisor recognises.

Ranking matters as much as the corpus. If results are ordered by citation count, a famous general review outranks the small recent study that measured exactly your question in exactly your country — and the famous one fills the space. Relevance has to come first, with citations as the tie-break.

A generated paper is a draft with a bibliography, not a finished piece of work. Read every reference you keep. Check the numbers that carry your argument. Delete the sections that say nothing — a tool asked for eight sections will produce eight.

Most institutions now require disclosure of AI assistance, and the requirement is usually about how it was used rather than whether. Say what it drafted and what you verified. That is a much easier sentence to write when the tool can tell you which claims it could not support.