While the term “autofiction” has been in use in French since the 1970s, it has only recently become common in English (Dix 2018). Its introduction has been contested; more than a few have suggested that the concept is modish, vacuous, or both (e.g. Caplan 2021). But what is the event of autofiction in English? Why did English come to require this term, and what has it been used to describe? Our project seeks to model an emergent “literary system” (Bode 2018)—insofar as this is possible with an inchoate and disputed genre category.

First we gather a corpus of autofiction in English (n = 195). For text selection we rely on “genre ascriptions”: explicit statements that a work is autofictional, whether found in academic discourse or literary discourse, on social media or the open web. Our corpus comprises works consistently described as autofiction across these domains, leading us to develop an “ascription corpus” of ~12k items. Corpus construction procedures like these, which have also been used to explore the “feminist novel” (Mandelman and Mukamal 2021) might well be adopted more widely in literary studies, especially where genre boundaries are shifting or new categories are coming into view.

Then, selecting only works published since 2000, we conduct a multifaceted comparison with the CONLIT dataset (Piper 2022), which presents derived data from ~2500 works of contemporary literature. Like CONLIT, we rely on BookNLP, a BERT-based annotation tool that tags text features including characters, places, events, and token supersenses (Bamman 2021). As Mélanie-Becquet et al. (2025) have recently shown in introducing BookNLP-Fr, a French-language adaptation of the BookNLP pipeline, BookNLP tagging can address subgenre classification problems like ours, which are of standing interest in computational literary studies (Allison et al. 2011; Hettinger et al 2016; Underwood 2016).

Autofiction has often been understood as a “hybrid” genre. So it presents an interesting test case for commonly-employed genre classification methods. We show that by using SVM on a set of Wordnet supersenses, autofiction can be distinguished from nonfiction (accuracy = 89.5%) and from fiction (82.6%). These results remain stable across feature ablations. Then, we project autofiction into the feature space defined by CONLIT fiction and nonfiction. The results show that autofiction clusters around the fiction-nonfiction hyperplane, while a majority of our texts (68%) lean toward fiction (Figure 1). The distribution of autofiction in this space, although wide, is unimodal. We note that unsupervised clustering methods on this group of texts tend to resolve into fiction and nonfiction, showing our autofiction as a subset of CONLIT fiction. Still, the supersense features that most strongly distinguish fiction and nonfiction (e.g. the verbs of perception discussed by Piper (2018)), are not the features that distinguish autofiction from fiction. On this definition anyway, English language autofiction is not a “hybrid” of fiction and nonfiction.

Figure 1: SVM Separation, Fiction versus nonfiction, autofiction projected.

Because many have claimed that autofiction represents an unfortunate turn inward, we turn to measures that express the shape and scale of narrated worlds. Where they emphasize internal reflection over external event, autofictions have often been described as “plotless” (e.g. Tanner 2022), such that some “autofictions” are also described as “autotheory” (Cavitch 2022). Indeed, autofictions turn out to have dramatically fewer “realis events” than do fictions (Sims et al. 2019). We also explored whether autofictional “navel gazing” might be operationalized using the measures of narrative “circuitousness” introduced by Toubia et al. (2021). This research shows that distances in vector space can describe disjunctions in a story’s time, setting, or perspective (Berger & Toubia 2024; Guest & Yan 2025). Piper and Toubia (2023) showed that higher “circuitousness” characterizes fiction when compared to nonfiction, again in CONLIT. Yet while its critics might suggest that autofiction is more ploddingly linear, conceptually slower, and imaginatively smaller than other kinds of fiction, we discover that these measures scarcely distinguish autofiction from nonfiction and cannot distinguish it from fiction. Turning finally to the consideration of character mobility and the “small worlds” of fiction quantified by Soni et al. (2023) and Wilkens et al. (2024), we report that autofiction’s protagonists appear in fewer geographical places than do the central figures of other kinds of fictional and nonfictional writing.

The event of autofiction is reflected in some, although not all, commonly-employed genre classification measures. The use of the term is not simply catachresis, nor does it simply denote a way of reading paratext, nor is it just a new means of selling books. Time will tell if this new species of expression will flourish in the evolving ecosystem of contemporary literature.