Other work on verifying citations, especially when a reference article is not in CrossRef, uses Google Scholar as a fallback (Topaz et al. 2026, Zhao et al. 2026). In testing VerusCite, I found enough errors that I intentionally avoided this approach.
Google Scholar, I believe, indexes items based on references it sees in papers. So if a paper has a citation that Google Scholar has not seen before, it is apt to actually create a new item for it. You can see how this can become a problem, as an article with hallucinated citations can actually create its own reference work.
For just one example, see this article in the Journal of Quantitative Criminology:
This was cited in a paper published in 2026 (that has now been retracted, but the original to view is in my benchmark data).
If you want further verification that this paper in the Journal of Quantitative Criminology does not exist, I am on the editorial board for JQC, and the paper is in my area of expertise. This paper decidedly does not exist.
I have found other similar indexing errors over time. So this is not an isolated example. Google Scholar actually indexes my personal blog, which is reasonable, as my personal blog has received a dozen or so citations (not from me!) over the years. So it is possible to not even publish in a peer-reviewed journal and trick Google Scholar into thinking a paper exists.
Perpetual bibliography errors
But one type of error I have noticed, which deserves slightly more attention, are bibliography errors perpetuated over time by Google Scholar. I have found a few interesting examples recently.
One example is a seminal citation in the recidivism literature on the risk-need-responsivity model:
So if you click on Google Scholar to export this, it cites a fictitious journal, Rehabilitation. Even includes an issue number!
Bonta, James, and Donald A. Andrews. “Risk-need-responsivity model for offender assessment and rehabilitation.” Rehabilitation 6.1 (2007): 1-22.
This however is a Canadian government report. You could argue the series is “Corrections research, user report 2007-06.” There is no Rehabilitation periodical with volume 6, issue 1, however.
So VerusCite will say this is a minor error. My instructions are to be more lenient with non-peer-reviewed work, as it can often be genuinely confusing how to cite reports that are not in traditional periodicals or books.
A second example: citing lme4
A second example I recently came across is citing a non-existent paper for the R lme4 package (note VerusCite gives the correct note; the R package itself suggests citing an article in the Journal of Statistical Software).
Bates, D. et al. Package ‘lme4’. Convergence 12, 2 (2015)
This has not accumulated 2,000 citations, but it does have a few, including some pre-dating ChatGPT. (So these are very unlikely to be “AI” generated hallucinations.)
I actually emailed Ben Bolker to confirm whether this periodical exists or not. He confirmed it does not, so like the Bonta example, this is an error that lives on, not necessarily because of AI, but because of people copy-pasting wrong information.
I have managed to
cite the correct version
of lme4. It is pretty easy, you just type citation(package) in R to find the correct
results.
VerusCite will not ever be able to find a smoking gun. Here it could have been that a user did a strange search in Google Scholar and copy-pasted the results. However, it is a pretty esoteric combination to get the false Convergence article to appear.
So whether it is laziness or an AI hallucination, I am not sure. I am leaving that flagged as a hallucination though in the Nature Human Behaviour article. Someone hallucinated it at some point, and Google Scholar’s error has made it stick.