Someone sent me an AI-drafted literature review last month. Twenty-three references. Clean formatting, no missing DOIs, authors and years and page ranges all in place. I resolved the first DOI out of curiosity. It didn't exist.
Neither did the second or third. Of the six that opened, two had the right title and author but argued something the draft never claimed.
That's worth writing down, because correct-looking formatting is exactly what a hallucination is good at. The model doesn't invent "some study". It invents a study with a plausible journal, a plausible volume, a DOI with the right prefix.
Fabrication comes in three grades
One: the reference is entirely invented. Author names that sound real, a journal that exists, a DOI with valid syntax. Checking takes ten seconds.
Two: the reference is real, the claim is not. This is the dangerous one. You search the title, you find it, the author and year match, so you trust it. But you're trusting the model's paraphrase, and the original might say "a weak positive association under specific conditions" while the paraphrase says "significantly improves".
Three: reference real, claim real, but your sentence overshoots it. Most common, and not really the model's fault. It happens when you rephrase: "improved" becomes "sharply improved", "some samples" becomes "generally".
Checking only grade one is what most people do. Grades two and three are where papers get rejected.
Three passes, in order
Pass one: does this reference exist?
If there's a DOI, paste it after https://doi.org/ and see if it resolves. No DOI? Search the exact title in quotes on Google Scholar or Crossref. Don't use fuzzy keywords.
A useful detail: hallucinated DOIs often have valid prefixes pointing at a real publisher with an invalid suffix. So don't judge by the prefix. Judge by whether it resolves.
Pass two: does the source say this?
Open the paper and search for the key phrase from your draft. If the phrase isn't there, the paraphrase was invented. No shortcut on this one.
If it's a PDF, read the paragraph around the hit, not just the hit. Most drift happens when a conditional becomes a statement.
Pass three: is my sentence as strong as the source?
Put the original sentence next to yours and compare word by word. The gap is usually in an adjective or a quantifier:
- source: "in some samples" → draft: "generally"
- source: "correlated with" → draft: "caused"
- source: "may" → draft: "confirms"
- source: "under condition X" → draft: "in general"
This pass is slow and it's the one that matters. Reviewers rarely reject over a nonexistent paper, because that takes one search. They reject because the citation can't carry the claim.
What software can and can't do
Be precise about the boundary. Inkory's AI check works on passes two and three. It compares your draft against the sources you actually stored — the highlights you made yourself — and flags where you wrote something stronger than the source supports, where the cited passage doesn't contain the claim, and where a sentence asserts something with no source behind it.
It does not do pass one, because it doesn't browse the web. That's a deliberate choice. A model asked to "verify facts" without access to them has one option: fill the gap. And what it fills in is formatted beautifully.
So the division of labour:
| Pass | Who |
|---|---|
| Does the reference exist | You (DOI resolution, exact-title search) |
| Does the source support the claim | Tool can flag it; you still open the source |
| Is the wording as strong as the source | Tool flags, you rewrite |
A procedure that saves drafts
If you have an AI-assisted draft right now, this takes about forty minutes:
- Export the reference list as plain text.
- Resolve every DOI. If one fails, delete the claim it supports. Don't "fix" the reference.
- For the rest, open each source and search for the key phrase.
- Where you find it, paste the original sentence next to yours and compare strength.
- Where you don't, downgrade the claim to "it is often argued" or cut it.
Step two deserves emphasis: don't replace a fabricated citation with a real one. When you do that, what you're really doing is finding a home for a sentence you already decided to keep. That's worse than an empty slot, because now you'll never look at it again.
The point
AI writing's real cost isn't bad prose. It's that it makes you feel like you have evidence.
The only defence is mechanical: every sentence with a citation should be one click away from the line in the source. Without that, a tidy bibliography is decoration.
Related: Getting Readwise highlights into your draft · 1,200 Obsidian clippings, twenty citations