1 in 277: Would your manuscript survive?
1 in 277 published manuscripts now carry a fabricated reference, and AI is the reason why. Would your last publication make the cut?
A new audit in The Lancet found fabricated citations in biomedical papers have risen twelvefold in two years, and the rise tracks almost exactly with when AI writing tools became widespread. The graph below shows the exponential climb. If you've used AI anywhere in drafting your last manuscript, this is worth your attention.
Why peer review isn't catching this
Reviewers are volunteers, working against deadlines, checking a manuscript's science rather than manually verifying every entry in a reference list. That's always been true. What's changed is the source of the errors. A misremembered citation from a rushed human author is a different problem to a fully-formed fake reference generated by a model that doesn't know the difference between recalling a fact and inventing one convincingly. Reviewers were never trained to catch the second kind, because until recently it didn't exist at this scale.
A hallucinated citation reads exactly like a real one. Plausible (or even a real) author, a real journal listed, the correct year. Nothing about it looks wrong until someone actually tries to find it.
Fake references are a real issue. Before AI, citation drift was already an issue; where the true finding of a source gets lost during citations across multiple manuscripts - because those citing it don't go back to check the original paper to confirm what was actually shown. Now, AI has brought this to a whole new level, where false references will be continually referenced in the future, misleading the knowledge base.
But thats not all.
If it's faking references, what else is it faking?
If a model will invent a citation with total confidence and no warning sign, what's it doing to the rest of your manuscript while you're not checking as closely? A summarised methodology that drifts from what was actually run. A statistic that's close enough to sound right but isn't what your data showed. Science already has a replication problem serious enough to have its own name. Feeding it references, and possibly data and methods, from a system that fabricates convincingly adds directly to that problem.
This is the position Co-Labb has held since I started this business out of my PhD.
When we've written manuscripts for teams at AstraZeneca and Genomics England, every citation, every figure, every methodological description was checked by a scientist who understood the underlying study, not a system that assembled something plausible and hoped it held.
I use AI in my own workflows. But, I don't let it near anything that goes out under a client's name unchecked, because I've read what happens when nobody does that checking.
What's actually on the line
A fake citation attached to your name is bad enough on its own. What's worse is what happens after someone finds it: a public correction notice sitting permanently against your paper, your name, and your organisation's name. Not so great for building trust and validating your science. A correction undoes the exact credibility a publication is meant to buy you, and it doesn't come off once it's there.
What to actually do about it
Don't wait for your journal's review process to catch this for you. Before your next submission, check every reference against a real source, not just the title and year, but the actual claim you're citing it for. Then check the data and the methods description against what actually happened in the lab. It takes time. It's tedious. It's also the only real insurance against your name sitting next to something that isn't true.
If you're going to publish a paper, do it right. Anything less isn't worth the risk to your name.
