Rettberg diagnoses LLM 'misaligned citations' as fabrication

Rettberg's paper, forthcoming in AI & Society, names two failure modes that make LLM-drafted academic prose look like scholarship without doing scholarship's work. 'Orbital argumentation,' a term she borrows from John Gallagher, is prose that circles a topic without landing on it. 'Misaligned citations' is her own coinage: the model reaches for the right author but the wrong paper.

Her running example is an AI & Society article by Robert Atkinson on digital temporality. Instead of citing Rob Kitchin's 2023 book *Digital Timescapes*, an obvious hit for the subject, the article reaches for an unrelated 2017 Kitchin paper. Judy Wajcman's *Pressed for Time*, another foundational work in the field, is absent. In a companion blog post, Rettberg puts it plainly: "Atkinson doesn't cite well-known research on his topic, but instead cites an irrelevant paper by the author of one of the books that should have been cited."

The mechanism, in her framing: "LLMs don't pick references logically. They pick them based on statistical pattern recognition, they pick them based on proximity between tokens in the text in the context window." The output is what she calls "drive-by citations, performative decorations that make a text look like scholarship without actually doing the work of scholarship."

The paper builds on Anne Sigrid Refsum's concept of floating motifs and the narratological idea of unnarratability, and treats these citation errors as a distinct form of scientific fabrication rather than ordinary sloppiness.

Rettberg also indicts her own venue. In the blog post she writes that "80% of the articles published in AI & Society seem to be AI slop these days." Two researchers we track on our Who's Who list had already shared the paper's link when we picked it up.