A Tale of Two Identities: An Ethical Audit of Human and AI-Crafted Personas
Pranav Narayanan Venkit, Jiayi Li, Yingfan Zhou, Sarah Rajtmajer, Shomir Wilson
TL;DR
The paper conducts an ethical audit of AI-generated personas, comparing 1512 synthetic narratives from three LLMs with 756 human self-descriptions to reveal how race-focused representations emerge in synthetic identities. Using close reading, TF-IDF, log-odds analysis, and a parameterized creativity framework, the study shows LLM personas disproportionately foreground racial markers, rely on reductive cultural tropes, and exhibit benevolent bias despite superficially positive narrations. It formalizes algorithmic othering as a lens on representational harms and demonstrates six harm categories (Stereotyping, Disparagement, Dehumanization, Erasure, Exoticism, Quality of Service) across synthetic outputs. The authors propose design recommendations for narrative-aware evaluation, human-centered validation, and transparency to mitigate harm in synthetic identity generation, arguing for authentic, community-informed approaches before deploying synthetic personas in sensitive domains.
Abstract
As LLMs (large language models) are increasingly used to generate synthetic personas particularly in data-limited domains such as health, privacy, and HCI, it becomes necessary to understand how these narratives represent identity, especially that of minority communities. In this paper, we audit synthetic personas generated by 3 LLMs (GPT4o, Gemini 1.5 Pro, Deepseek 2.5) through the lens of representational harm, focusing specifically on racial identity. Using a mixed methods approach combining close reading, lexical analysis, and a parameterized creativity framework, we compare 1512 LLM generated personas to human-authored responses. Our findings reveal that LLMs disproportionately foreground racial markers, overproduce culturally coded language, and construct personas that are syntactically elaborate yet narratively reductive. These patterns result in a range of sociotechnical harms, including stereotyping, exoticism, erasure, and benevolent bias, that are often obfuscated by superficially positive narrations. We formalize this phenomenon as algorithmic othering, where minoritized identities are rendered hypervisible but less authentic. Based on these findings, we offer design recommendations for narrative-aware evaluation metrics and community-centered validation protocols for synthetic identity generation.