AI Killed My Journal Submission

In April of 2023, I submitted a paper to a fairly well-known philosophy journal from a major publishing house. This was the second journal I’d submitted this paper to—it had been rejected from a different one years ago, with some helpful and generous comments, and I’d procrastinated sending it out again. I had a co-author for this paper, and we both spent time considering different places where our paper would be a good fit, until we agreed on one.

After a couple weeks of status limbo, during which our manuscript was “undergoing initial checking” (an ambiguous phrase that could mean almost anything), the paper was sent back to me with a cursory note:

I’m sorry to have to report that we are unable to use your paper. After careful consideration by our editors, we regret to inform you that we must decline this submission on editorial grounds and subsequently have declined to send the paper out to external peer reviewers. Our decision is based on our automated plagiarism report of your submission running at 42%.  We can see that a bulk of your submission is directly quoted from another publication of yours, entitled [REDACTED]. We like to ensure at [REDACTED] that the submissions are completely original and are not published or under consideration elsewhere, even if it be in parts.

This was startling to me. The email was largely correct—large portions of our co-authored paper (although not nearly 42%…) were taken from the aforementioned manuscript of mine. But the email was incorrect in one crucial aspect: that manuscript of mine was unpublished. It was a chapter of my dissertation, and I had been required to upload my dissertation to an online database once I had successfully defended it at the end of my PhD program. The paper I was currently trying to publish was a project that combined parts of my work from that dissertation chapter with work that my co-author was doing. None of the “plagiarized” sections of the co-written work were taken from anything published.

It was clear from the email that this was an entirely automated process. I doubt any human had actually set eyes on the details of the report beyond seeing “42%” and thinking, “Well, that seems pretty bad” and hitting the “reject” button. I figured that a simple email might clear things up and allow the paper to go back under consideration (or at least make it past the AI plagiarism guards). But emails to both the handling editor and the editor-in-chief went unanswered. Ultimately, we ended up just submitting the paper to a different journal, where it is currently under review.

The Kafkaesque (if I may use the term here) nature of this experience is far from unique. AI use in the realm of academic publishing is keeping pace with the general uptake of AI in every other aspect of society, which is to say, at an incredible rate. The Chronicles of Higher Education recently ran a long piece on the experience of academics with AI. Although ChatGPT has gotten the most attention of all programs at the intersection of academic writing and AI, it is far from the only player in the game. As Taylor Swaak writes, after extensive interviews with academics on their use of AI tools in paper reviewing:

There’s Explainpaper, where one can upload a paper, highlight a confusing portion of the text, and get a more reader-friendly synopsis. There’s jenni, which can help discern if a paper is missing relevant existing research. There’s Quivr, where the user can upload a paper and pose queries like: What are the gaps in this study?

In their book The Phantom Pattern Problem, Gary Smith and Jay Cordes, two data scientists, detail a serious epistemic predicament of our current culture of seemingly-infinite data. They call this the “phantom pattern problem”. In short, given the eye-watering surplus of data available to us (and exponentially more available to big-data algorithms), we are able to uncover and identify far more patterns in that data than ever before. And while this is certainly helpful and informative (else the algorithms would be of no use to us!), it is not the case that all patterns in data are meaningful. Some patterns are mere coincidences, and the more data you have, the more meaningless patterns will arise. These are what Smith and Cordes call “phantom patterns”, which they define as “coincidental correlations [that] are useless for making predictions” (2020, p. 45). In 2022, I published a paper on phantom patterns in big data (Fritts and Cabrera 2021) with the same co-author. Fitting, somehow, that only the next year we should fall victim to this very phenomenon!

It is common to experience the effects of AI identifying “phantom” patterns, usually with little consequence to us (perhaps our credit card is frozen after we shop at an out-of-state grocery store). In my case, it is easy to see why an algorithm would flag my submission as plagiarism: paragraphs copied nearly verbatim from a different paper located in a publicly-available database. The details and nuances (variety of “publication”, the distinction between dissertation databases and journal databases, copyright issues, differences between verbal similarity and plagiarism) that make my sort of case interesting are difficult to teach to a basic algorithm of the kind that was checking for plagiarism in my case.

It is in situations like this, in our daily interactions with new technology, that we find ourselves praising what we might call a “human touch”. “If one single human had looked at our submission before letting the robots reject it,” I complained to my co-author, “it would be under review by now!” While humans are certainly susceptible to the “phantom patterns problem”, we are also better at identifying instances of it, because we are uniquely capable of understanding the difference between meaningful and non-meaningful correlations. We are uniquely capable of understanding meaningfulness.

If it is the recognition of the meaningful that human analysis alone can contribute to data analysis, then it is also this feature that gets left behind in areas where human labor is displaced by AI. In another paper of ours (Fritts and Cabrera 2022), we explore the charge that AI “dehumanizes” the realms in which it replaces human laborers. We argue that the “dehumanization” in question here refers to the ways in which AI possesses “artificial values” (to borrow a term from Nguyen 2021) given to it by its creator, rather than the complex and opaque “real values” possessed by humans. It is this “dehumanized” aspect of AI in the workforce that drives our discomfort at the idea of, for instance, using AI chatbots as crisis hotline counselors. The difference between “real” and “artificial” values can be seen clearly in the case of my plagiarism charge. But the utility of using AI in academic publishing is equally obvious, especially as the peer-review crisis works its way up to a breaking point. Finding solutions to this crisis has seemed a monumental task, with journals financially unable to pay reviewers without charging unaffordable fees for paper submissions. AI seems perfectly poised to provide an easy and equitable solution. And perhaps solving this problem is worth an occasional false charge of plagiarism. But to what extent would we, or should we, accept the artificial values of AI over real human values in academic publishing? Will we trust AI to tell us when a paper has cited an insufficient number of sources? Should we rely on machine-learning algorithms to help us comprehend a particularly dense paragraph of a manuscript we are reviewing? Are there aspects of academic publishing and reviewing that require the real, complex, opaque values of human minds? Arguably, these are questions that academic publishing should focus on sooner rather than later.

Megan Fritts is Assistant Professor of Philosophy at the University of Arkansas at Little Rock. She works in several subfields of philosophy, including philosophy of technology, action theory, epistemology, and applied ethics. Megan is also the co-host of the podcast Philosophy on the Fringes.