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

Introduction

One evening in early 2024, in a small control room at a Chinese AI startup, I watched the debut of a generative-AI-powered female avatar on a major Chinese e-commerce platform as she promoted a luxury skincare line (see Figure 1). She was not modeled on any single person. Instead, developers assembled her from traits that market data suggested would sell: balanced facial symmetry, a warm vocal timbre, and micro-expressions calibrated to appear inviting. Yet this final product was not the outcome of a single design decision, but of continual adjustments. At launch, she spoke in a calm, authoritative tone, explaining the science behind the ingredients. Within minutes, viewers began clicking away, the chat went silent, and sales flat-lined. Behind the monitors, developers tracked performance dashboards in real time, then softened her voice, inserted gentle compliments, and animated her into a warmer smile. By midnight, watch time had doubled and sales rebounded. “We found her parameters,” one engineer said. What I witnessed captures the dynamic of this study: femininity was treated as an instrument, rendered as measurable, adjustable variables, and re-assembled until it aligned with commercial goals.

OPEN IN VIEWER

Figure 1.

Screenshot of an AI streamer in a livestreaming event.

AI products have been widely critiqued for reproducing narrow, feminized defaults, from voice assistants to recommendation systems (Buolamwini and Gebru, 2018; Woods, 2018). Much of this critique follows a bias-reproduction account: biased inputs yield biased outputs. This framework has been valuable for diagnosing gender bias, but it treats the production process as a black box. We know what goes in and what comes out, but not what happens in between. Inside that black box, contemporary AI systems involve complex production pipelines where bias can enter at many points: corpus curation, alignment with human feedback, decoding settings, and online A/B experimentation (Ho et al., 2025; Ouyang et al., 2022). We know a lot about why biased inputs produce biased outputs, but far less about how gender is actively manufactured through this iterative process.

This article opens that black box by examining how gender is produced through what I call the optimization loop: a recursive, metric-driven cycle in which user feedback is used as a design lever to generate, refine, and scale the variant that best meets commercial goals. Drawing on sociological accounts of gender as something “done” rather than possessed (Butler, 1990; West and Zimmerman, 1987), I describe this process as “doing data, doing gender.” Where traditional “doing gender” depends on interpersonal recognition to validate performances, “doing data” shifts validation to commercial metrics that select which performances survive. Developers begin with culturally familiar stereotypes and translate them into quantifiable parameters. Those parameters are then tested against commercial metrics, which select and intensify configurations based on what converts. The resulting avatars do not reflect what real women are, nor pre-existing audience ideals. They reflect what the optimization loop learned would make audiences stay, click, and buy. Audiences participate in this loop not as judges but as generators of behavioral data: their watch time, clicks, and purchases become the feedback that shapes the personas engineered to capture their attention.

Drawing on 120 h of ethnographic observation inside two Chinese AI startups and 48 interviews with developers, designers, and brand marketers, my study opens the black box of AI production to trace how femininity is created and stabilized through two pivotal phases of the optimization loop. In the pre-launch stage, developers translate vague brand ideals such as “spicy” or “sweet” into measurable parameters like voice pitch, facial softness, gaze patterns, and gesture timing. Although presented as neutral technical features, these translations encode culturally normative assumptions about femininity, laying the groundwork for commercial appeal. Once the avatar is launched, each parameter enters a recursive A/B testing process governed by platform-specific metrics such as watch time, click-through rate, and sales per minute. Traits that consistently boost these metrics are retained and often codified into “reference avatars,” while underperforming variants disappear. Across repeated cycles, the “winning” style emerges as a hegemonic femininity (Schippers, 2007): soothing, admiring, deferential. It appears natural only because it reliably converts attention into profit.

By shifting the analytic lens from bias reproduction to commercial-metric production, this article advances scholarship on gendered AI in three ways. First, it reframes AI systems not as static products that mirror culture, but as dynamic production sites where gender is constructed, stabilized, and scaled through iterative optimization. Second, it shows how developers use feedback as a design lever, reverse-engineering “winning” variants and sidelining others under platform metrics and commercial goals. Third, it shows how big-data infrastructures (A/B testing stacks, metrics dashboards, traffic-allocation systems, and write-backs) obscure human choices in the production pipeline, so the narrowing of gendered personae appears as objective, data-driven discovery. Together, these insights move AI-ethics debates beyond bias detection toward the infrastructural and commercial processes through which identities are produced, commodified, and standardized.

Gender in AI: From bias reproduction to productive optimization

Research on gender bias in AI converges on a bias-reproduction account: biased inputs yield biased outputs. When texts link “man” with programmer and “woman” with homemaker, models reproduce those links (Bolukbasi et al., 2016; Caliskan et al., 2017). In vision, when training sets and labeling schemes underrepresent darker-skinned women or collapse gender into a binary, systems make more errors on darker-skinned women and misrecognize or erase non-normative genders (Buolamwini and Gebru, 2018; Keyes, 2018; Scheuerman et al., 2019). Complementing data-centric critiques, research on designer positionality shows that who builds systems, and in what organizational contexts, shapes problem framing, data choices, and evaluation criteria (Cambo and Gergle, 2022; Leavy, 2018; Miceli et al., 2020; Scheuerman and Brubaker, 2024). At deployment, feedback loops in recommendation systems amplify what already draws clicks, shaping what users see (Abdollahpouri et al., 2019; Chaney et al., 2018; Singh and Joachims, 2018). Across these strands, bias is understood as something that enters the system and is reproduced or amplified at the output.

At the interface level, a growing body of work examines how gender is designed into AI products. Popular voice assistants such as Siri, Alexa, and Cortana are given feminized names, voices, and personas, positioning them in service roles that reinforce patriarchal gender norms (Dobrosovestnova et al., 2022; Leavy, 2018; Woods, 2018). Strengers and Kennedy (2021) show how “smart wives” are designed to be helpful, compliant, and emotionally available, embedding gendered assumptions about domestic labor into the interface. These interactions often follow sexualized scripts, with users issuing commands and assistants responding with deference (Costa and Ribas, 2019; Hester, 2017; Schiller and McMahon, 2019). This work reveals how designers pre-script gender into AI products. Yet it focuses primarily on design intentions and interface features, treating gender as something coded once and deployed. What remains underexplored is the iterative process through which gendered personas are tested, refined, and recalibrated after launch.

This article shifts the focus from bias reproduction to gender production. Existing literature diagnoses bias by examining inputs (skewed data, designer assumptions) and outputs (unequal error rates, stereotypical personas). But the production process itself remains a black box: we know what goes in and what comes out, but not how gender is actively manufactured in between. My study opens this black box ethnographically by tracing what I call the optimization loop: the recursive, metric-driven process through which developers produce gendered AI. In this process, audience feedback is not just observed but operationalized as a control signal that shapes design. Developers, designers, and marketers use behavioral metrics to select, refine, and scale gendered traits. Crucially, developers do not invent femininity from scratch; they begin with culturally familiar stereotypes such as youthfulness, softness, and deference. What the optimization loop adds is a mechanism for filtering and intensifying particular configurations based on metric performance. Traits survive not because they are culturally dominant but because they convert. By tracing this pipeline, I locate human agency throughout: in the choice of what to measure, what thresholds to set, what counts as success, and how results are written back into production.

The optimization loop works as follows. Developers translate brand adjectives into tunable traits such as voice pitch, gaze, and gesture (Scheuerman et al., 2021), link them to dashboards, and run A/B tests. Metrics such as watch time, click-through, and sales-per-minute are treated as ground truth, and traits are reverse-engineered from these numbers until they maximize performance. High-performing variants are filtered through ranking and recommendation systems (Bucher, 2018; Gillespie, 2014; Rieder and Hofmann, 2020) and scaled as defaults, while weak ones are dropped. Over repeated cycles, platforms allocate traffic to the “winning” traits, narrowing the field toward a single profitable version of femininity. Gender is not simply carried forward by existing bias but actively assembled through a metric-driven workflow that turns femininity into a design object optimized for commercial goals (Wajcman, 2010).

This productive perspective builds on theories of gender as an ongoing accomplishment rather than a fixed essence. West and Zimmerman's (1987) notion of “doing gender” and Butler's (1990) account of performativity highlight how gender is stabilized through the repetition of stylized acts such as tones, expressions, and gestures that others recognize as coherent. Traditionally, this recognition is interpersonal: social audiences validate performances as appropriately masculine or feminine. I argue that commercial AI introduces a new mode of doing gender, which I call “doing data.” Here, gender is still accomplished through repeated performances (e.g. scripted voices, facial animations, gestural cues). What shifts is how performances are validated. Rather than social audiences judging appropriateness, commercial metrics such as watch time, click-through rates, and sales conversions determine which performances survive. In an optimization loop, gender is first rendered into measurable parameters, yet the choice of what to parameterize encodes cultural assumptions into ostensibly objective data (Kitchin, 2014). Those parameters are then tested against metrics, which select configurations based on what converts. “Doing data” is thus a commercially mediated way of doing gender, where measurement encodes norms and metrics select which configurations are standardized and scaled.

Audiences participate in this process not as judges of appropriateness but as generators of behavioral data. Their watch time, clicks, and purchases become the signals that validate which performances survive. Critical algorithm studies have shown how platforms construct audience identities by correlating their behavioral traces into categories (Bucher, 2018; Chun, 2021; Gillespie, 2016; Wang and Spronk, 2023). My study shows that the same behavioral traces also flow in the other direction: they become design inputs that shape the personas audiences encounter. Audience data not only classifies who is watching but also engineers what is being watched. In Cheney-Lippold's (2017: 5) terms, the avatar becomes what it is “computationally calculated to be.” Thus, audience feedback is not merely a downstream readout but a design lever that developers use to iteratively select, refine, and scale gendered traits. Understood this way, gender in commercial AI is not a fixed attribute encoded once and deployed, but a moving target continuously recalibrated through the optimization loop.

The Chinese case: Gendered AI streamers in live-commerce

Live-commerce refers to real-time video shopping with in-stream checkout, where viewers watch a streamer demonstrate products and purchase without leaving the stream. What distinguishes live-commerce from other live retail formats, such as television shopping or traditional e-commerce like Amazon, is its metrics-driven architecture. Streamers do not perform for an imagined audience; they perform for the metrics displayed on dashboards in front of them. Platform algorithms allocate traffic based on real-time performance indicators such as watch time, click-through rates, and sales per minute. Streamers who generate strong metrics receive more traffic; those who underperform see their streams deprioritized. This algorithmic distribution of visibility creates intense pressure to optimize every aspect of the performance: voice, gesture, pacing, and emotional tone are all calibrated to move the numbers. It is this metrics-driven logic that has shaped the rapid adoption of AI streamers, whose parameters can be adjusted in real time to maximize commercial performance.

In China, live-commerce is not a niche experiment but a mainstream platform economy operating at a national scale. Major platforms, including Taobao (Alibaba), Douyin (ByteDance), JD, and WeChat Channels, have integrated livestream shopping into their core business models. By the end of 2023, 816 million people—about three-quarters of Chinese internet users—had participated in livestream shopping, and industry revenues surpassed US $700 billion, with projections to exceed US $1 trillion by 2026 (Statista, 2024). Live-commerce thus stands as one of the largest big-data-mediated consumer economies in the world, producing massive, continuous streams of behavioral data that shape not only consumer experiences but also the organization of labor.

Operating livestreams at this scale requires considerable resources. Human streamers command high labor costs, tire over long sessions, and differ in their ability to deliver the affective style preferred by brands. These constraints have fueled the rapid adoption of AI-generated “digital humans” (shuzi ren 数字人): photorealistic avatars created using generative AI, deepfake technologies, and lip-sync models. Major technology companies such as Baidu, Tencent, and ByteDance have developed the underlying AI infrastructure, while a growing ecosystem of smaller production companies specializes in creating customized digital humans for brands. By the end of 2024, approximately 1.36 million companies in China had registered in the “digital human” category, a 36.9% increase from the previous year (China University of Communications, 2025). The digital-human market was valued at US $26 billion in 2022, with the majority of applications concentrated in live-commerce (iiMedia Research, 2023). My fieldwork takes place inside two production companies that create and optimize AI streamers for brand clients.

The rapid rise of AI-generated “digital humans” in Chinese live-commerce offers a powerful site for studying gendered AI. Live-commerce in China is highly gendered: beauty and fashion, categories targeted primarily toward women, account for over 70% of sales, and roughly 80% of human streamers are young women trained to project warmth, charm, and relational engagement. The Chinese term for streamers, zhubo (主播), carries gendered connotations in this context. Female streamers in China's digital economy often face sexualized stigma, and fashion and beauty influencers (wanghong网红) navigate similar gendered expectations (Craig et al., 2021). These existing cultural scripts provide the templates that brands and AI companies draw upon when designing AI avatars. AI streamers overwhelmingly reproduce this female-coded persona: they are typically designed as youthful, attractive women whose appearance and demeanor draw on deepfake models, lip-sync animation, and generative AI dialogue systems. Both startups I studied produced almost exclusively female avatars. Male avatars exist primarily in categories such as electronics and tech. However, developers explained that female personas “convert better” across most product categories. Even campaigns targeting female audiences used female avatars. As one product manager put it, “Women trust women for beauty advice, and men like watching women. Female just works.” Male avatars occasionally came up in discussions but were dismissed as “niche” without proven metrics.

Yet these personas are not fixed. Traits such as voice pitch, gaze angle, facial softness, and micro-expressions are parameterized and adjusted in collaboration between vendors and brands, often during livestreaming sessions. Viewer metrics such as watch time, emoji bursts, and add-to-cart rates serve as yardsticks: underperforming variants are quickly tweaked or discarded, while successful versions are cloned and scaled across storefronts. This data-intensive process can unfold in just days, with enormous commercial impact: AI avatars in China have sold millions of dollars’ worth of goods in single livestream sessions (Cheng, 2025). This case provides a fertile ground for analyzing how gender in AI is not merely represented but actively produced through iterative cycles of optimization and metric-based feedback within big data infrastructures.

Data and methods

This study draws on fieldwork at two AI startups in Hangzhou, a city often described as China's “digital capital” for its dense cluster of AI startups specializing in live-commerce technologies. The two “digital human” production companies I worked with—pseudonymized here as startup A and startup B—supply avatars to brands livestreaming across China's major live-commerce platforms (Taobao, Douyin, JD, and WeChat) and collectively hold a significant share of the national AI-streamer market. I conducted approximately 120 h of participant observation inside these two companies, following full AI-streamer production cycles from initial client brief to post-launch analytics. This involved shadowing engineering teams on the sales floor, sitting in design meetings where voice, face, and gesture parameters were debated, and observing real-time “war-room” sessions during 12 live broadcasts. During these sessions, I recorded how developers altered avatar traits mid-stream in response to shifts in performance metrics such as watch time, emoji bursts, and add-to-cart rates. Fieldnotes were written within 24 h of each visit and supplemented with screenshots of metrics dashboards, photographs of whiteboard sketches, and internal presentation slides shared by staff.

To triangulate my findings, I also conducted systematic online observation of 30 AI-streamed sessions across major platforms. These observations extended my participant observation beyond the production sites, allowing me to compare the internal design decisions I witnessed with their public enactment. They provided a comparative lens on how gendered performances were staged, evaluated, and monetized in live streams, and offered a check on whether the practices I observed inside companies aligned with broader market patterns.

Alongside observation, I conducted 48 semi-structured interviews lasting between 60 and 90 min with engineers (n = 9), photographers (n = 3), avatar designers (n = 8), product and project managers (n = 15), and brand-side marketers (n = 13). The design teams at both startups were predominantly male, particularly in engineering and product management roles, while marketing and brand liaison positions included more women. The workplace culture reflected this composition. Developers rarely used gendered language to describe their products, emphasizing instead technical sophistication, smoothness of animation, realism, and customization options. Metrics were treated as ground truth, and engineers followed standardized workflows organized around performance dashboards. Gender, when it surfaced, was framed as a product variable to be optimized rather than a social category to be discussed. Interviews were conducted in Mandarin, audio-recorded with informed consent, transcribed verbatim, and translated where necessary.

Whereas participant observation allowed me to witness the micro-level design practices, interviews provided access to the rationales participants themselves used to justify or contest these design choices. Questions focused on the design workflow, the role of performance metrics, client-vendor negotiations, the perceived effectiveness of gender cues, and the challenges of maintaining “authenticity” in AI-mediated sales. Comparing how practitioners narrated their decision-making with what I observed in practice revealed both alignments and tensions between discourse and action. These interviews were further complemented by the analysis of 73 documents, including internal materials (KPI reports, marketing decks, technical proposals) and industry reports. These documents provided a check on interview accounts and revealed the formal evaluative criteria used to assess avatar performance.

Data analysis followed an abductive coding strategy (Timmermans and Tavory, 2012), combining open coding of fieldnotes, interview transcripts, and documents in MAXQDA with iterative memo-writing to refine emerging conceptual categories. Codes such as “parameter translation,” “metric spike,” and “variant cloning” were progressively clustered into higher-order themes that mapped onto the optimization loop framework. This abductive approach allowed me to move between empirical detail and theoretical insight, sharpening the account of how gendered traits were produced, evaluated, and stabilized in practice.

My position as a Mandarin-speaking researcher with prior industry familiarity facilitated access to design spaces and informal conversations. Although the research centers on two large Hangzhou-based firms representative of mainstream, revenue-driven production practices, it provides a rare, fine-grained account of how gender is operationalized, tested, and monetized within China's live-commerce AI industry.

Findings

This section traces how gender is actively produced and stabilized through an optimization loop spanning the entire development cycle of AI streamers. I organize the findings into two stages that correspond to distinct phases of production. Stage 1 (pre-launch translation) follows a sportswear brand collaboration at startup A to show how abstract brand aesthetics are systematically translated into contractually fixed parameters before any audience interaction occurs. Stage 2 (live optimization) draws on multiple campaigns at startup B to examine how these parameters enter recursive testing cycles driven by real-time platform metrics. Across repeated cycles, high-performing variants are scaled as defaults while underperforming ones are dropped, progressively narrowing the design space toward a single profitable femininity.

Stage 1: Pre-launch translation: From vague vibes to metric-ready parameters

This stage unfolds in three substages. First, developers translate vague brand ideals into measurable parameters. Second, they assemble the avatar by compositing the face and voice from source materials. Third, iterative negotiation with clients locks the parameters into a final persona ready for launch.

Stage 2: Post-launch optimization: From engineered parameters to hegemonic femininity

With parameters locked into contracts, the avatar was ready for launch. This stage unfolds in three substages. First, developers tune avatar parameters based on real-time performance data. Second, the repetition of winning traits trains audiences to expect and reward them. Third, winning configurations are codified and scaled across the industry, hardening into hegemonic femininity. This section draws primarily on fieldwork at startup B. I analyze a skincare campaign that illustrates avatar tuning through A/B testing and a whisky brand collaboration that reveals how audiences are trained to respond to stabilized affective cues.

Discussion and conclusions

This study examines how commercial AI systems manufacture gendered personae through a recursive, metric-driven optimization loop. Drawing on ethnographic fieldwork at two Chinese AI startups and 48 interviews with engineers, designers, and marketers, I show that AI streamers become “feminine” not through a one-time design choice but through continuous calibration to real-time performance data. Brand descriptors such as “sweet-spicy” or “approachable” are first translated into tunable traits such as voice pitch, gaze curvature, waist-to-hip ratio, and scripted emotional tone, then formalized as production assets (prompts-of-record, reference faces, design tokens, contracts). After launch, these traits are A/B-tested in livestreams. Platform and business metrics such as watch time, emoji reactions, add-to-cart rates, and sales per minute determine which combinations are retained. High-performing configurations are codified as standardized “reference avatars” and carried forward in backlogs and budgets, while underperforming variants are dropped. Across cycles, metrics functions as a control signal for design, narrowing a wide space of possible styles to a single, profitable, platform-legible femininity: soft-spoken, admiring, emotionally available. Femininity is not merely reflected; it is assembled, selected, and stabilized to maximize commercial goals.

This production focus extends the bias-reproduction account that dominates scholarship on gendered AI. Existing work shows how skewed data, labels, developer demography, and ranking dynamics yield gendered outputs (Abdollahpouri et al., 2019; Buolamwini and Gebru, 2018; Chaney et al., 2018; Leavy, 2018; Singh and Joachims, 2018). These accounts treat bias as something that enters the system and is reproduced or amplified at the output. The optimization-loop view shifts the focus from reproduction to production. Feedback is not merely a downstream readout of existing bias but a design tool that actively shapes what the AI becomes. The loop links four moves into a single production system: translating gendered brand adjectives into controllable parameters; linking those parameters to performance metrics; reallocating exposure toward high-performing variants; and writing winners back into production artifacts. In this sequence, data do not merely register a pre-given persona; they select and stabilize a gendered one.

Because each step is framed as “what performs,” the politics of gender are naturalized as neutral performance engineering. Developers of AI streamers seldom aim to “add femininity”; they chase performance metrics and adjust traits that fail to lift the numbers. Yet this purported neutrality reliably rewards a narrow cluster of affects (softness, emotional availability, an admiring tone) while penalizing frictional styles. Across thousands of small, data-driven tweaks, rewarded traits harden into a template that later reads as natural and “user-driven.” The engineered identity is then attributed to “what the data wants,” producing infrastructural plausible deniability: hegemonic gender identity becomes hard to contest precisely because it appears as the market's objective will rather than the cumulative effect of metric-guided selection by human teams. This diagnosis locates agency not in data alone but in the human choices that define metrics, thresholds, allocation rules, and write-backs.

The optimization loop contributes to theories of identity in AI by treating identity as actively produced rather than merely reflected. Critical data studies show how platforms manufacture user identities by correlating behavioral traces and sorting audiences (Bucher, 2018; Cheney-Lippold, 2017; Chun, 2021; Gillespie, 2014). I extend this claim to the supplier side: the same infrastructures of correlation, ranking, and reward also construct the AI persona itself. Building on “doing gender” (West and Zimmerman, 1987) and performativity (Butler, 1990), I argue that optimization infrastructures function as an engine of metric recognition. It is performance against metrics, not interpersonal recognition, that decides which stylized acts “count.” Repetition under reward sediments high-performing acts into a stable persona, which is then written back into production plans. Gender is thus not a fixed attribute traveling through a pipeline; it is the moving result of parameterization, measurement, selection, and codification.

This understanding of AI identity as produced rather than reflected has political-economic implications. The optimization loop is the routine product-work through which surveillance capitalism acts on identity (Zuboff, 2019). In this economic model, platforms do not simply predict what people will do; they intervene to shape those actions. Every pause, scroll, click, or purchase becomes behavioral data, which feeds back into algorithms that determine which design features will maximize engagement (Fourcade and Johns, 2020). In this model, AI avatars are not merely outputs; they are instruments for generating behavioral data. Commercial AI avatars are engineered not just to represent gender but to elicit the attention, clicks, and purchases that platforms extract and monetize. Gender itself becomes a variable that can be tuned, tested, and monetized, continuously recalibrated to increase “behavioral surplus”—the excess attention, engagement, and purchasing extracted once the system identifies what works. When a particular look, tone, or movement improves the metrics, it is reinforced in the next design cycle until audiences are trained to respond to that version. The optimization loop applies the extractive logic of surveillance capitalism to gender, narrowing the range of possible performances and naturalizing the most profitable form as inevitable.

These dynamics extend beyond commercial AI to carry implications for gender inequality in the wider social context. The traits that consistently “work” (softness, deference, emotional availability) are precisely those that align with patriarchal expectations: femininity that legitimates male authority and provides emotional labor without reciprocity (Connell, 1987; Schippers, 2007). What makes this process powerful is that commercial viability and gender hierarchy converge on the same traits. Because selection occurs through metrics rather than explicit ideology, this convergence is masked as neutral market logic. As these avatars proliferate across platforms, they normalize this narrowed femininity as “what works,” potentially shaping expectations for human women in similar roles. The optimization loop does not just carry existing gender inequalities forward; it amplifies them, scales them, and presents them as what the market naturally rewards.

Making the optimization loop visible re-politicizes gender in AI by relocating accountability from “upstream hygiene” to the optimization infrastructures that decide which personae get scaled. Instead of treating bias as a residue of skewed inputs, this framework shows how identity is actively produced through metric-driven optimization, what I call doing data, doing gender. The accountability question turns on authorship of metrics and thresholds, rationales for exposure decisions, and the organizational routines that construct and stabilize winning traits. Gendered outcomes are not the impersonal verdict of markets but the cumulative record of situated choices by developers, product managers, and platform operators. This opens new terrain for analysis and governance: assessing decision rights over metrics, tracing selection histories, and evaluating how “metric recognition” substitutes for interpersonal recognition in making identities count.

This study situates its analysis in China's hyper-commercial live-commerce sector, a magnified site where the optimization loop is unusually exposed at platform scale. Here, AI avatars are valued only insofar as they move measurable outcomes, and performances are continuously scored by retention, reactions, and add-to-cart behavior. Under these high-frequency feedback conditions, the commercial calibration of gender becomes visible: femininity is not coded once and for all but continually recomputed at the pace of analytics. Focusing on this site expands gender-and-AI research beyond Euro-American voice assistants such as Siri and Alexa (Noble, 2018; West et al., 2019) and surfaces dynamics that remain harder to see in less commercial or less transparent settings.

The mechanisms traced here are not unique to live-commerce. Iterative A/B testing, metric-guided elimination of alternative styles, and the survival of the most commercially performant persona occur wherever AI products test affective cues against performance metrics. Voice assistants, customer-service bots, social companions, or educational avatars are all subject to similar optimization pressures. Early evidence points to domain-specific affective norms: efficiency-oriented styles in customer-service chatbots (Følstad and Skjuve, 2019); relational attachment and empathy cues in social robots (Darling, 2016); and supportive, empathic behaviors in educational robotics (Belpaeme et al., 2018). Comparative work can trace optimization loops across domains and countries to identify which performance indicators crystallize which hegemonic styles, and how those styles intersect with race, age, or class. Metric-driven femininity in Chinese live-commerce is thus not an outlier but a diagnostic lens on how big-data optimization infrastructures worldwide assemble normative personae in AI systems.