Virginia Conde

Fifteen years of experience in human neurophysiology and cognitive neuroscience 🧠 1st-gen academic

Current position: Senior Advisor at the Centre for Digital Scholarship, University of Oslo Library

Trying to understand where we're collectively heading but, most importantly, why 🗺️

pronouns: she/her
blog: https://critical-overthinking.ghost.io

Last Notes

2026-07-18 09:49:31 UTC

- reply

yes, precisely; there is an issue with the order of events, which at the very least should be iterative -- I write a draft of something / an idea / a guess, wonder if someone has looked into this, search for studies, read/skim articles, inform my idea/guess, reshape my writing.
The alternative is to write something I believe / guess / reason, but because I'm writing an *academic article*, I need to make it look like one, so I plaster it with references that seemingly support my assertions (but that I haven't read).
A lot of this discussion seems to circle around whether the references are fake or whether they do exist. But a reference "existing" is not enough if the author hasn't engaged with the text in any capacity, or if it hasn't informed their thinking and reasoning in any way. At this point it ceases to be an academic approach and turns into mere aesthetics.
It is very easy for a GPT model to provide a reference as an output to a prompt, many of which do exist, while still misattributing information that is not supported by said reference. The only way to ascertain whether a reference has been correctly attributed is to know what the article contains; but if you already know the reference and have engaged with its contents, you wouldn't really need for a GPT model to do that work for you. If you need to check all references for accuracy and relevance anyway, then how does using a GPT model (instead of, say, Scopus, or Web of science) save you time?

2026-07-18 07:51:31 UTC

- reply

this is definitely a problem in and of itself; particularly if the references were only added *after the fact* (the writing) and not used as an a priori to support claims or assertions.
Part of the problem is that *even if the references do exist* ("the links go were they're supposed to") there's still a big risk of misattribution and misreferencing when following this approach.
I've tested this myself with my own article to see if GPT4o and GPT5o would refer to my study through a relevant prompt; in one case, it didn't refer to the study at all (even though it has 250+ citations to date, and it's from 2019), in the other, it referred to the study but got the methods/approach wrong (I later found out that those methods belonged to a different publication, not mine).
Long-story short, if all the references were known *a priori* by the author and it was a case of mere retrieval, I could see how it'd be less of a problem (still weird to use GPT for this given we already have accurate search engines for this purpose).
If the author was using GPT to *search* for references they were not already familiar with, and doing this a posteriori once they had already written the piece (so, an a posteriori add-on of study references to support claims already made), then it is a big, big issue.