How to Check the Citations an AI Gives You

One of the most dangerous things a large language model does is produce citations that look completely real and are entirely invented. The author names sound plausible, the journal title fits the topic, the year is reasonable, and the whole thing is formatted correctly. None of that means the source exists. Treating AI-supplied references as trustworthy is how fabricated citations end up in real essays and reports, and it is a mistake that can sink your credibility instantly. This guide walks through how to check every one.

Why models fabricate references

A language model does not have a card catalog it looks things up in. It predicts text that is statistically likely given your prompt. A citation is just a very patterned piece of text, so the model is extremely good at generating something that has the shape of a citation without any guarantee that it points to a real document. It may blend two real papers into one, attach a real author to a title they never wrote, or invent a plausible page range and DOI out of thin air. These are not rare glitches; they are a predictable consequence of how the technology works. Assume every reference is guilty until proven innocent, no matter how confident, specific, or well-formatted it looks — confidence and specificity are exactly what a model is best at producing, real source or not.

Why this happens: a quick mechanism check

It helps to understand, even briefly, why this failure is so consistent across different tools and different topics. A language model generates text by predicting a plausible continuation, token by token, based on patterns it absorbed during training — it does not maintain or consult a database of verified sources the way a library catalog does. A citation is, to the model, a small and highly patterned piece of text: author, year, title, venue, page range. It is very good at producing something with the right shape. Nothing in that process checks whether the specific combination corresponds to a real document. This is covered in more depth in a companion piece on why AI fabricates citations in the first place, but the short version is enough to justify the discipline in this article: treat shape and truth as two separate properties, because the model only guarantees the first one.

Verify the source actually exists

Start with existence before you worry about accuracy. Work through each reference in order:

1. Search the exact title in quotation marks in a scholarly database or a general search engine. A real paper almost always surfaces immediately.
2. Resolve the DOI if one was given. A valid DOI links to a real landing page; a fabricated one will fail to resolve or lead somewhere unrelated.
3. Look up the author and confirm they work in this field and have published on this topic. A mismatch is a red flag.
4. Check the venue — confirm the journal or publisher is real and that a volume and issue for that year actually exist.

If you cannot find the source through any of these routes, do not use it. A citation you cannot locate is not a citation; it is a liability, and no amount of confident phrasing around it in your own text changes that.

Verify the source says what the AI claims

Existence is only half the job. A real paper can still be misrepresented. The model may cite a genuine study but summarize a conclusion it never reached, or invert a finding, or borrow a number from a different table. Once you have the actual document open, read the relevant section yourself, slowly, and confirm three separate things: that the claim is actually present somewhere in the text, that it means what the AI said it means once you read it in its original context, and that any figure or quote attributed to it is reproduced exactly rather than paraphrased into something slightly different. Pay special attention to quoted material and page numbers, which models fabricate readily. If you cannot find the sentence you are supposedly citing, the citation does not support your point, no matter how real the paper is.

Verification takes real time — plan for it

It is worth budgeting for this honestly rather than treating "just check it" as a free action squeezed in at the last minute. A responsible existence-and-relevance check on a source — enough reading to confirm it exists, that the author is credible, and that the abstract plausibly supports the claim — runs to a few hundred words of reading per source. Across a reading list of a dozen AI-suggested sources, that adds up to a genuine chunk of time, easily half an hour or more before you have read a single source in full. Treat citation verification the way you would treat any other identifiable phase of a project: give it a real slot in your schedule, sized honestly using whatever reading-time estimate fits your actual source list, rather than assuming it will happen for free in the margins of drafting.

A field-by-field verification log

A lightweight log turns "I'll remember to check that" into something you can actually track and revisit. For each AI-supplied reference, record: the exact title as given, whether a search for that exact title returned a real matching document, whether a provided DOI actually resolves, whether the named author works in a plausibly related field, and a short note on whether the specific claim attributed to the source was confirmed by reading the relevant passage yourself. A reference that has not cleared all five checks does not go in your bibliography yet, and the log tells you at a glance which ones are still outstanding rather than relying on memory during a busy final editing pass.

Build a habit that scales

Verification feels tedious the first few times and then becomes fast. A few habits make it sustainable:

• Never let an unverified reference sit in your document, even temporarily. Mark it clearly, for example with a bracketed note, so it cannot slip through.
• Log the source the instant you confirm it — capture the DOI, the access date, and the exact page — so you never have to re-find it.
• Prefer sources you can read in full over anything you can only see in a snippet.
• When the AI gives you a claim without a source, treat that as a research to-do, not as evidence.

A hypothetical, walked through step by step

Concretely, imagine an AI assistant supports a claim about group decision-making with a reference formatted like: Whitfield, K. & Anzalone, M. (2020). “Consensus Drift in Small Deliberative Groups.” Journal of Behavioral Group Dynamics, 8(2), 144–161. This is a deliberately invented example for illustration — do not search for it, it does not correspond to a real publication. Notice, again, how unremarkable it looks; that is the entire mechanism of fabrication, not a special case. Working through the four-step check: a title search for the exact phrase turns up nothing matching; the named journal either does not exist or, if a similarly-named one does, has no volume 8 matching this description; neither author is easily confirmed as working in this specific area; and there is no DOI offered to resolve at all, which is itself a signal to slow down rather than a neutral omission. Any one of those four failures is sufficient reason to drop the reference. Seeing all four together is exactly what a fabricated citation typically produces once you actually look, however convincing it read on the page.

Preprints, published versions, and which one you actually read

One verification detail that catches even careful writers: many papers exist in more than one version — an early preprint posted before peer review, and a later published version that can differ in its results, its framing, or even its conclusions after revision. An AI-supplied citation may point to the wrong version of a genuinely real paper, or blend details from both without distinguishing them, and a title search alone will not surface this discrepancy since both versions typically share the same title and authors. When a specific number or claim matters to your argument, confirm which version you are actually citing — the preprint or the peer-reviewed version — and cite that exact one, since a claim that changed between versions is a real source of error even when every other part of the citation check passes cleanly.

What to do when a citation fails

When a reference turns out to be fabricated, resist the temptation to ask the same model to “fix” it or supply a replacement, because it will very often just invent a fresh fake with the same confident tone as the first one. Instead, take the underlying claim and go find real support for it yourself, using a library database, a reference manager, or the bibliography of a paper you already trust. If no real source supports the claim, then the claim itself may be wrong, and you have just been saved from publishing an error.

The rule is short and worth memorizing: a citation from an AI is a lead to investigate, never a source to cite. Do the checking, keep the receipts, and your bibliography will hold up to any scrutiny it faces.

Once every reference in your list has actually been verified this way, the remaining work is mostly mechanical — organizing what you have confirmed, formatting it consistently, and keeping it in sync as your draft changes. That mechanical layer is covered separately, since it is a genuinely different set of habits from the verification work described here, and it is worth treating the two as sequential rather than trying to format a reference before you have finished confirming it is real.