A large-scale computational analysis of 13,869 fiction novels published between 1485 and 2025 indicates that an author’s gender does not consistently predict whether female major characters are depicted in gender-stereotypical ways. The investigation, conducted by researchers Svenja Kratzke, Erdem Meral, and Hannes Rosenbusch and published in EPJ Data Science, evaluated character portrayals using public book records and reader perceptions extracted across decades of literary output. While past scholarship often suggested that male writers depict female characters in substantially more stereotypical manners than female writers do, the new findings demonstrate that the observed differences are largely attributable to the confounding influence of literary genre and publication year. When statistical models controlled for the genre in which an author writes, the broad main effect of author gender on character stereotypicality disappeared. Instead, the specific traditions of narrative genres and the broader historical era proved to be vastly stronger determinants of how female protagonists are characterized in popular fiction.
The study was designed to address a fundamental methodological question in literary text analysis: whether previously documented gender gaps in narrative characterization reflect genuine authorial biases or simply the uneven distribution of male and female authors across different literary genres. Prior investigations examining historical fiction samples noted that male writers often framed female figures with lower agency, relegated them to recipients of actions rather than active agents, and described them with a stronger visual emphasis on physical attractiveness. Theoretical models such as Social Dominance Orientation have frequently been cited to argue that male authors possess greater motivation to reinforce traditional hierarchies through narrative conventions. However, genre demographics present a classic confounding variable known as Simpson’s Paradox. For example, female writers comprise approximately 95 percent of the romance genre, which inherently foregrounds interpersonal warmth and domestic communion, whereas male writers historically accounted for around 70 percent of science fiction, an arena characterized by active problem-solving and adventurous agency.
To rigorously disentangle author identity from genre conventions, the research team assembled a broad corpus derived from a collection of public Goodreads reading logs initially compiled by Rosenbusch and Korthals. The original source collection contained 49,674 unique titles identified through a Python script that randomly sampled Goodreads user profiles and gathered their reviewed books. From this broader repository, the researchers isolated 17,363 novels and novellas categorized explicitly as fiction. To filter and annotate this voluminous dataset, the investigators employed a search-enabled artificial intelligence agent based on the Perplexity Sonar model operating at temperature zero. The agent searched online public repositories, encyclopedia articles, reviews, and catalog records to determine whether each book prominently featured a female major character. Titles lacking female leads or containing unresolved author metadata were systematically excluded, yielding a final analytical sample of 13,869 English-language books with fully documented author genders, publication dates, and genre classifications spanning more than five centuries.
The annotation pipeline eliminated several problematic data categories
The annotation pipeline eliminated several problematic data categories before statistical modeling commenced. The researchers removed 101 exact duplicate records, 180 instances where automated generation broke off prematurely, 340 books lacking sufficient online documentation for reliable annotation, 57 cases where author gender was non-binary or unidentifiable, and 192 books with indeterminate publication years. In a manual cross-validation of 100 author gender estimates within thriller and crime genres, the researchers identified occasional complexities, including pseudonym shifts, post-publication name changes, and collaborative series continuations under single legacy names. These nuances indicated that approximately 5 percent of author attributions in broad corpora carry inherent classification ambiguities. The AI agent categorized each book into primary genres, identifying Romance, Fantasy, Historical Fiction, Mystery, Literary Fiction, Young Adult, Thriller, Science Fiction, Horror, and Children’s Literature as the ten most prominent categories, with unclassifiable works assigned to a residual category.
Character portrayals were systematically evaluated using thirteen newly developed annotation item pairs grounded in established social psychology frameworks, particularly the Stereotype Content Model. This psychological model conceptualizes social perception along orthogonal dimensions of competence or agency and warmth or communion. The study’s items measured specific narrative manifestations: agency and competence items assessed whether protagonists solved challenges through skill, took risks, required rescue by male characters, or occupied submissive roles; warmth and communion items assessed caretaking, domestic roles, and emotional attunement; and additional items tracked occupational rank and physical appearance focus. Each item pair featured independent stereotypical and counter-stereotypical formulations, allowing annotators to record nuanced portrayals where a character might display both traits simultaneously. Net stereotypicality was calculated by standardizing stereotypical and counter-stereotypical scores against non-missing responses and subtracting the counter-stereotypical score from the stereotypical score, generating a continuous metric where positive values represent predominantly stereotypical portrayals.
To establish the reliability of automated annotations extracted from secondary sources, the researchers conducted detailed human-in-the-loop validation experiments. A manual audit of twenty randomly sampled novels, encompassing 580 individual judgments, compared the AI agent’s outputs directly against cited source texts. This verification revealed annotation accuracy rates ranging between 80 percent and 100 percent across items, with complete agreement on objective metadata such as author gender, publication year, and overarching genre. Annotation discrepancies primarily stemmed from rhetorical nuances, such as ironic reader reviews, or ambiguous phrasing in plot summaries. A subsequent validation exercise evaluated 100 items on a seven-point continuous scale using a second independent human reviewer and an alternative web-enabled language model, GPT-5.1. The ratings from the second human annotator and GPT-5.1 demonstrated a strong positive correlation of r = 0.85, confirming that missing values and labeling variations reflected genuine information sparsity in online documentation rather than idiosyncratic model hallucinations.
Statistical modeling of the full dataset revealed
Statistical modeling of the full dataset revealed that initial apparent differences between male and female authors disappeared when genre interactions were introduced. In a basic linear regression without interaction terms, author gender exhibited a statistically significant association with net stereotypicality, with books authored by men scoring slightly lower in overall stereotypicality by an average difference of -0.28 standard deviations. However, when an interaction term between author gender and literary genre was added to the model, the main effect of author gender ceased to be statistically significant, yielding a t-statistic of 1.839 with p = 0.066. In contrast, the author-by-genre interaction term proved highly significant (F = 177.63, p < 0.001), demonstrating that the relationship between author gender and characterization varies directionally across different literary categories rather than operating as a universal, uniform bias across fiction.
Genre classification emerged as an exceptionally strong predictor of character portrayal across all tested specifications (F = 101.284, p < 0.001). Historical Fiction, Romance, and Literary Fiction consistently featured the most stereotypical female major characters in the corpus, whereas Science Fiction, Mystery, Thriller, and Fantasy consistently displayed the most counter-stereotypical portrayals. Within specific genres, post-hoc comparisons showed multi-directional patterns: male authors authored significantly less stereotypical portrayals than female authors in Fantasy, Romance, and Science Fiction, whereas female authors wrote significantly more counter-stereotypical characters in Romance and Mystery. Concurrently, publication year exhibited a robust, highly significant negative relationship with net stereotypicality (b = -0.0055, p < 0.001). Female major characters have steadily gained agency and shedding stereotypical constraints over time, a trend that remained fully robust when the authors conducted sensitivity analyses excluding historical outliers published before 1975.
When isolating individual social perception dimensions, author gender remained largely uninformative, with two distinct exceptions. For the broad competence and agency dimension, secondary sources indicated that female authors created characters perceived as slightly less competent than those created by male authors, though the effect size was negligible (eta-squared = 0.0001). The single domain displaying a substantial and consistent author gender disparity was physical appearance focus. When evaluating whether a female protagonist was explicitly admired for her physical beauty, male authors scored significantly higher than female authors across nearly all genres (mean difference = 0.28, d = 0.29, p < 0.001). This specific finding corroborates earlier linguistic research on the literary male gaze, indicating that while male authors do not necessarily diminish their female characters’ competence or agency, they remain substantially more likely to foreground physical attractiveness in narrative descriptions.
The study authors emphasize several critical methodological boundaries
The study authors emphasize several critical methodological boundaries inherent to their reliance on secondary online documentation. Because the automated pipeline parsed Goodreads summaries, Wikipedia articles, and book reviews rather than full primary narrative texts, the findings capture public character reception and widely documented traits rather than subtle, fine-grained syntactic structures. This focus inherently limits the corpus to books with sufficient online documentation, leaving under-documented indie literature and non-Western titles less represented. Furthermore, the dataset was constrained to books featuring female major characters, meaning it does not capture stereotyping that occurs through the wholesale exclusion of women from central narrative roles. Nevertheless, across nearly fourteen thousand titles, the overarching evidence demonstrates that cultural genre expectations and historical evolution exert a far more decisive influence on character depictions than the gender of the individual holding the pen.
Subject of Research: Computational analysis of female character stereotypes across literary genres and publication history
Article Title: Do female and male authors differ in their portrayal of female characters? An AI-assisted exploration of public book records across genres and time
Article References: Kratzke, S., Meral, E., & Rosenbusch,
Article References: Kratzke, S., Meral, E., & Rosenbusch, H. (2026). Do female and male authors differ in their portrayal of female characters? An AI-assisted exploration of public book records across genres and time. EPJ Data Science, 15(1), Article 66. https://doi.org/10.1140/epjds/s13688-026-00665-y
DOI:10.1140/epjds/s13688-026-00665-y
Keywords: literary stereotypes, computational humanities, female major characters, author gender, genre analysis, book history, stereotyping trends, natural language processing, Goodreads corpus, character agency, warmth and competence, literary fiction, science fiction, romance novels, AI annotators, EPJ Data Science, character representation
News. Literary Genre And Era Outweigh Author Gender In Book Character Stereotypes. Scienmag
