Katherine Elkins Identifies Three Structural Limits on AI Fiction

Training on vast libraries of human fiction seems like the obvious path to teaching machines to write it. Yet a paper by Katherine Elkins, presented at the MFS Cultural AI Conference at Purdue University, argues the opposite: the more fiction AI systems have absorbed, the more visible their structural inability to generate it becomes.

Elkins identifies three architectural reasons for this. First, compelling narrative requires events that feel surprising in the moment yet retrospectively inevitable -- a temporal quality that conflicts directly with how transformer-based models generate text forward. Second, fiction regularly asks readers to reweight the significance of details after the fact, but current attention mechanisms have no way to perform that kind of retrospective revaluation. Third, effective fiction coordinates sentiment across multiple scales simultaneously -- individual words, sentences, scenes, and the full emotional arc -- in a way that exceeds what models can currently manage.

The paper draws on over seven years of collaborative research on sentiment arcs, and Elkins frames the diagnosis not as a temporary capability gap but as a theoretically precise account of why the architecture itself resists what great fiction demands. Whether that framing holds is the part the paper does not settle: diagnosing structural limits is not the same as proving they cannot be engineered around, and no proposed architectural fix is offered here.

The stakes are oddly double-edged. Elkins notes that AI companies have risked billion-dollar lawsuits for access to modern fiction -- a bet the architectural argument suggests may not pay off creatively. And in the other direction, the paper flags that genuine fiction mastery by AI would enable human manipulation at scale, a risk that, if current limitations are ever overcome, would arrive without obvious governance frameworks to contain it.

For now, the honest read of this work is that creative AI tools are better understood as sophisticated drafting aids than as authors. The unresolved question is whether these three limitations are ceilings or just today's floor.