Study: AI as an Author Simplifies Identity into Simple Classes

American researchers have examined large language models as cultural agents and authors: how they produce and limit perceptions of authorship and identity in the current American literary field.

The researchers created a simulation featuring 101 "AI authors." These characters were modeled after real authors, and generative AI was asked to write texts that mimic their style and background. This allowed for an examination of how the language model constructs literary distinction, for example, in terms of race, gender, and publication context.

According to the analysis, the literary world produced by AI heavily emphasizes a simple black-and-white division. This dichotomy flattens both the internal diversity of groups and excludes other possible axes of distinction, such as various cultural and social backgrounds. AI-generated texts were also compared to those written by humans, highlighting the systems' tendency to homogenize expression and identities.

Based on the results, the model treats identity primarily as a limited, categorical variable – not as a living and evolving cultural practice. In other words, identity functions for AI as a label that guides the use of predefined features and clusters, not as a multi-layered narrative or experience.

The study does not focus on how well AI imitates an individual author, but rather on what kind of understanding of the literary field and its power relations the model produces in general. The observation of AI's homogenizing, simplifying way of handling identities raises the question of what kind of image of literature and its actors generative models convey to their users.

Source: The social AI author: modeling creativity and distinction in simulated cultural fields, AI & SOCIETY.