Academic publishing contains a specific kind of irony. While the rest of the field advances, a researcher spends months or even years condensing truly helpful ideas into a structured paper, submitting it to a reputable journal, getting it accepted, and then watching it remain mostly unaltered in a digital archive. One of the most prominent publications in computing, the Journal of the ACM, has experienced its fair share of these silent disappearances. But every now and then, something changes. A forgotten paper reappears because the world finally realized what it was saying, not because anyone went looking for it.
With a 2019 study that hardly registered when it was first published, that appears to be happening right now. In its first two years of publication, the paper, which dealt with classification systems, algorithmic behavior, and the edge conditions under which machine reasoning fails, received only a few citations. Enough to imply that it had been seen. Not enough to indicate that it had actually been read. Scholars in related fields looked past it. It lacked the attention-grabbing hook that draws readers in a crowded publication cycle.
The paper did not change. It was everything in its immediate vicinity.
Scholars began working backward as AI safety research accelerated in 2023 and sharply increased through 2025, looking for theoretical underpinnings in areas that had been missed during the initial excitement of the large language model boom. Safety researchers, many of them trained in machine learning rather than classical computer science, were encountering problems that felt new but weren’t entirely. The question of what a model truly knows versus what it seems to know, the behavior of systems under distribution shift, and the dependability of outputs in adversarial situations were all well-known issues. They were associated with older papers, older frameworks, and older names.
It turns out that some of those questions had been raised by the 2019 ACM study before the current generation of models existed to make them urgent. That timing is almost uncomfortable. Before the field had the vocabulary or institutional pressure to take it seriously, the paper was published. Both are now present, and citations are arriving in groups.
Although this pattern isn’t specific to this paper, it does have special significance in light of the broader issues surrounding research integrity. According to a recent Lancet study, between 2023 and 2025, the number of fabricated citations—which point to papers that don’t actually exist and are frequently attributed to AI tools hallucinating references—rose sixfold. Once a sort of silent agreement between the writer and the reader, the bibliography is now a contentious area. In light of this, it almost feels like a correction to see a real paper finally receive genuine engagement. It’s not very dramatic. However, a correction.
It’s difficult to ignore the slight difference between the automated noise that researchers are now trained to distrust and the current surge of citations surrounding this paper. Its actual arguments, not just its title, are typically discussed in the papers that cite it. Compared to five years ago, that distinction is more significant today. The issue with AI-generated citations, according to researchers like Maxim Topaz at Columbia, whose work tracking citation fraud has garnered widespread attention, is not only inaccuracy but also the creation of a false sense of a field’s history by covering up actual intellectual lineages with fictitious ones.
In some respects, the 2019 ACM paper is the reverse of that issue. It’s a genuine intellectual heritage that was just ignored for too long. The work was completed, published, and put away. The issues it brought up remained. They held out.
It’s still unclear if this increased focus will actually have an impact on the advancement of AI safety research rather than just showing up in reference lists. Engagement and citation are not always synonymous. Watching this develop gradually across preprints and journal issues, however, gives the impression that the field is doing something it doesn’t always manage: carefully examining the past before proceeding.
