Why I Let GPT Write My Related Work Section (And You Should Too)

I want to be honest with you. I have not read all the papers I have cited.

I have read all the titles. I have read most of the abstracts. I have read the introductions of the ones that seemed relevant and the conclusions of the ones that seemed long. I have read the methodology sections of approximately twelve papers in my entire career, and those were papers I was reviewing, and two of them I did not finish because the notation was unreasonably dense and I had a 5 PM deadline.

I am telling you this because I believe you have done the same thing, and I think we should discuss it.

The Economics of Citation

A citation serves two purposes. The first is to acknowledge prior work. The second, and let us be honest about this, is to demonstrate that you are aware of prior work, which is not the same as having read it. The demonstration is what matters. The awareness is theoretical.

Citations are a signaling mechanism operating in a system where the signal has decoupled from what it purports to signal. We cite papers to indicate engagement with the literature. The literature is now so large that full engagement is not possible. We have therefore developed a set of citation practices that gestures toward engagement while permitting the actual engagement to be selective, strategic, and sometimes entirely absent.

The Citation Farm

I have cited papers because the reviewer asked me to. I have cited papers because they were cited in papers I was citing, and adding them made my related work section appear more comprehensive. I have cited papers by researchers I hoped would review my work favorably, on the grounds that people generally review positively work that cites them, and I say this with confidence because I have experienced this from both sides.

I have cited papers that, upon later inspection, do not say what I implied they said. Not because I was dishonest — I would like to be clear about that — but because I read the abstract, inferred the argument, and was wrong. The paper argued something adjacent to what I needed it to argue. I cited it anyway. The reviewers did not check.

A Defense

The citation norm exists because the alternative is worse. A field with no citations would have no record of its own development, no way to trace the lineage of an idea, no mechanism for credit allocation. The citation system, with all its dysfunction, is better than no citation system, in the same way that peer review is better than no peer review, even on the days when you open Reviewer 2’s comments and consider a career in farming.

So I will continue to cite. I will continue to cite things I have not fully read, because I cannot fully read everything, because there is too much of everything, because the field produces knowledge faster than any individual can consume it. This is the condition of contemporary academic publishing. I did not design it. I am simply, like everyone else, trying to navigate it while appearing to have read the Bengio 2013 paper that everyone cites and almost no one has read in full.

You know the one.

The author declares the following citations were inserted at the specific request of Reviewer 2: [1], [3], [7], [12], and [14]. References [4] and [9] were inserted preemptively, for the same reason.