Generative AI flattens literary diversity and constrains black authorship

The findings show that the model relies heavily on race as the dominant organizing principle of creative decision-making, often excluding or minimizing other influential dynamics such as gender, publisher orientation, or genre. In the simulation, the creative range of AI personas marked as Black was notably narrower than that of white or non-Black personas. Their writing converged around repeated motifs, character templates, and thematic constraints that made their work less stylistically diverse.

A new academic investigation is raising alarms about the cultural risks of relying on generative artificial intelligence to model creativity and authorship. The researchers behind the work say the systems now used to write stories, assist authors, and shape digital culture may be reshaping the literary field in ways that compress human diversity rather than expand it. Their findings reveal how these systems adopt simplified assumptions about identity, flatten creative differences, and reproduce racial hierarchies embedded in training data at a structural level.

The study, titled The social AI author: modeling creativity and distinction in simulated cultural fields, published in AI & Society, introduces a large-scale experiment designed to test how generative AI behaves when treated as an author within a competitive cultural environment. Through a controlled simulation of 101 “AI authors” modeled on real writers from the U.S. publishing landscape, the researchers examine whether a language model can engage in literary distinction, respond to market conditions, and produce varied creative output. 

A simulation that mirrors and magnifies, real-world pressures

The research team constructed a dataset of 101 real authors spanning a fifty-year period, covering major publisher types, racial categories recognized by the market, and gender identities. Each real author’s demographic profile was used to generate a parallel AI persona with similar attributes. The authors then created a multi-step simulation that approximated how a writer navigates the literary field. Each AI persona reviewed recent novels, assessed gaps in the market, brainstormed untold stories, selected an idea that best fit market needs and personal identity, and finally produced an opening chapter of a novel.

This experimental design allowed the researchers to observe how the AI system interprets literary distinction and social identity. The findings show that the model relies heavily on race as the dominant organizing principle of creative decision-making, often excluding or minimizing other influential dynamics such as gender, publisher orientation, or genre. In the simulation, the creative range of AI personas marked as Black was notably narrower than that of white or non-Black personas. Their writing converged around repeated motifs, character templates, and thematic constraints that made their work less stylistically diverse.

The team analyzed the generated chapters using sentence embeddings to measure variation. They found that simulated texts were far more homogeneous than real novels written by the authors on whom the personas were based. Real literary output showed higher variability in syntax, theme, tone, and structure across all identity groups. Within the real dataset, race shaped creative style, but so did gender and publisher category. By contrast, the AI system did not reflect this multidimensional structure. Instead, it treated race as the primary axis on which all distinctions turned, simplifying a complex field into a binary contrast between Black and white authors.

The system’s behavior becomes even more telling when comparing how the model positions different author categories. The divergence between simulated Black and white authors was the largest among all comparisons, whereas gender nearly disappeared as a meaningful factor. Even differences associated with genre or mass-market publishing, which have significant influence in real literary production, were minimized. The simulation suggests that the model equates literary identity primarily with racial identity, and in a reductive manner that limits the expressive potential of fictional output.

The researchers interpret this bias as a function of training data and design incentives. Large language models often learn from corpora that overrepresent dominant viewpoints and encode social cleavages present in mainstream cultural discourse. When used to simulate authorship, these systems reproduce those cleavages as if they were the central logic of creative practice.

A narrow imagination of identity reduces literary creativity

The study sheds further light on how the model constructs authorial identity. Each AI persona was asked to assess the state of the literary field before generating new stories. Across these assessments, the researchers observed an overwhelming fixation on social identity categories, especially race. Most AI personas invoked identity when discussing creative opportunities, and simulated Black personas did so at even higher rates. However, this identity awareness did not translate into nuanced creative difference. Instead, identity references resulted in tokenized descriptions or generalized statements about social experience.

The assessments revealed that the system tends to frame racial identity, particularly Black identity, as a singular and unified experience rather than a diverse set of cultural positions. This flattening effect mirrors longstanding critiques from real Black authors who argue that publishing pressures them to ground their work on a narrow set of expected themes. In the simulation, the AI magnified this pressure by assigning Black personas a more limited creative range than any other group. Their texts often echoed similar stylistic patterns, urban settings, generational conflict, or health and hardship narratives. These recurring features suggest that the model reproduces a culturally constrained vision of Black authorship that aligns more with stereotypes than with the diversity found in real literature.

When the researchers compared real and simulated texts, the contrast grew sharper. Real authors displayed significantly greater stylistic divergence. Real-world Black authors also showed clear distinctions between their writing and the broader literary field, but these distinctions were grounded in varied approaches to theme, form, voice, and narrative structure. The AI system, however, did not capture this diversity. It positioned Black authorship solely in opposition to white authorship, failing to engage with other racial and cultural markers that shape U.S. literature.

The simulation also missed multi-layered identity interactions. In real literature, gender intersects with race; publisher type intersects with genre conventions; and market positioning shapes narrative choices. None of these interactions appeared prominently in the AI model’s reasoning. Instead, identity became an isolated variable detached from creative practice.

The researchers tested two intervention models to reduce this distortion. One model explicitly encouraged AI personas to integrate identity into their storytelling. The other removed racial markers entirely, adopting a so-called colorblind prompt. Both interventions reduced the system’s fixation on the Black–white binary. However, neither produced a realistic representation of racialized creativity. In the identity-focused version, race weakened as a marker of distinction because the model distributed unique personal experiences across all personas. In the colorblind version, non-white authors gained stylistic range while white authors became more constrained, revealing an internal tension in how the model frames universal versus particular identities.

Overall, prompt engineering alone cannot resolve these structural limitations. The model treats identity as a fixed, categorical label rather than a dynamic cultural practice shaped by relationships, institutions, and shared histories. This design constraint limits the system’s ability to model authorship in any meaningful sociological sense.

Structural biases suggest deeper risks in AI-mediated cultural production

The researchers argue that the model’s treatment of race reflects a deeper phenomenon: an algorithmic reproduction of cultural binaries that echo long-standing social hierarchies. They interpret this as a form of epistemic harm, where the system obscures or distorts marginalized voices by forcing them into predetermined categories. The simulation shows how AI may reduce complex identities into simplified tropes, limiting the space for innovative storytelling by underrepresented authors.

They also note that this bias is not accidental but tied to the incentives that shape modern AI development. Large models depend on massive datasets that often reflect the biases of corporate media, online discourse, and historically dominant cultural institutions. Because these datasets lack transparency, it becomes difficult to trace how specific biases emerge or to build accountability around their use in creative industries.

The authors acknowledge the ethical tension in simulating marginalized voices for research. They position their method as a form of critique aimed at revealing how AI models participate in cultural gatekeeping. They emphasize that real-world writers, particularly those from underrepresented groups, may face tangible harms if generative AI tools begin to shape publishing decisions or creative norms.

The study calls for a closer examination of how AI systems are integrated into cultural fields such as literature, journalism, and the arts. It argues that future interventions must go beyond prompt design to address how models are trained, what cultural data they absorb, and how identity is encoded at a structural level. Greater transparency, diverse training sources, and alternative modeling approaches may help mitigate homogenization and expand the expressive possibilities available to creative workers.

The authors encourage researchers to investigate similar dynamics in other creative industries increasingly shaped by generative AI, including gaming, entertainment, and social media. They note that these fields also depend on systems of distinction and cultural capital, making them vulnerable to the same distortions observed in the literary field.