Doing data, doing gender: Manufacturing gendered AI through the optimization loop

Abstract

Scholars increasingly warn that commercial AI products reproduce narrow, stereotypical gender identities, but far less is known about how those identities are made in practice. This article addresses that gap through the case of AI streamers in China's expanding live-commerce sector, where generative-AI streamers are built to appear hyper-feminine and sell products as stand-ins for human streamers. Drawing on 120 h of behind-the-scenes ethnography in two Chinese AI startups and 48 interviews with engineers, designers, and brand marketers, I show that an AI streamer's gender is produced, not merely reflected, through an optimization loop: a recursive, metric-driven cycle in which developers generate, refine, and scale the variant that meets commercial goals. Pre-launch, teams translate abstract brand ideals into parameters across voice, face, gaze, gesture, and script. Post-launch, continuous A/B tests link these parameters to performance metrics (retention, click-through rate, sales per minute). Exposure is reallocated to higher-performing variants, and the winners are written back into the product as defaults. Across cycles, data do not simply register a pre-given persona. They select and lock in a gendered one, yielding a soft-spoken femininity optimized for sales. This article extends bias-reproduction accounts by unpacking the production pipeline of AI products and showing that user feedback is not a mirror of preexisting bias but a design lever teams use to reverse-engineer persona traits. This reframing shifts accountability from “bad data” to human choices and makes clear that identities in AI are engineered, traceable, and therefore contestable.