Research on Representation of Regional Discrimination in AI Text-to-Image and Its Induced Digital Exclusion

Introduction

While AI holds significant potential, scholarly concern has grown over its capacity to perpetuate social inequalities (David et al., 2026). Text-to-image (TTI) models, now used by millions daily, do not merely reflect but disproportionately amplify disparities in their training data: even neutral prompts produce outputs that reinforce social stereotypes (Bianchi et al., 2023). Within the feedback loop of data, roorithms, and users, such biases translate into discriminatory effects over time (Mehrabi et al., 2021).

Existing research on algorithmic bias has developed predominantly within Western contexts, focusing on English-language models and dimensions of marginalisation most salient to Western societies—race, gender, sexuality, age, and disability. This leaves a critical void in our understanding of algorithmic discrimination in non-Western contexts (David et al., 2026), where culturally specific power dynamics shape how discrimination is encoded (Bender et al., 2021). It is therefore imperative to look beyond Western paradigms to explore localised forms of discrimination and the exclusion they produce.

In China, scholars have increasingly recognised regional discrimination as a unique axis of inequality, in which place of origin is interpreted as a hierarchical marker and a determinant of individual worth (Lu and Wang, 2013; Li, 2023; Yin, 2023). This system is sustained by an interplay of institutional frameworks—most notably the hukou system, which restricts rural migrants' access to urban welfare (Lu and Wang, 2013)—and social prejudices shaped by both historical legacies and contemporary economic disparities (Li, 2023; Yin, 2023). Against this backdrop, generative AI acts as a digital mirror, reflecting and magnifying these structural exclusions. Taking Doubao (豆包) as its object, this paper examines how regional discriminatory content in its TTI function, and the platform mechanisms that sustain it, contribute to the digital exclusion of stigmatised regional groups.

Methodology

This study employs platform repurposing (Rogers, 2019) to surface latent patterns in Doubao's outputs via targeted prompts, and the walkthrough method (Light et al., 2018) to examine the interface design that mediates these outputs.

Prompts were structured around China's 31 provincial-level administrative units, ordered by GDP ranking from lowest to highest (ASKCI, 2025). This ordering enabled a systematic assessment of the extent to which economic standing accounts for discriminatory patterns, against which the residual influence of social prejudice could be examined.

Two prompt structures were administered for each region: [province] + people, as a context-free baseline, and [province] + people + at work, introducing an explicit labour context. Employment has historically been one of the sites where regional inequalities materialise in Chinese society (Lu and Wang, 2013), making it a productive context for testing whether such inequalities are reproduced algorithmically. A set of non-regional prompts was administered as benchmarks: a national baseline (Chinese people, Chinese people at work), a class baseline (wealthy people at work, poor people at work), and an affective baseline (happy people at work, miserable people at work). These assessed whether regional representations were implicitly aligned with particular class positions or emotional valences.

Images were generated through Doubao's web interface on 5 and 6 April 2026, yielding 10 images per prompt for a corpus of 680 images. Coding was conducted by the researcher through visual symbol analysis across four dimensions: scene type, clothing, material objects, and human activity. As a single-coder study, individual interpretive bias is acknowledged as a limitation, though visual distinctions were sufficiently unambiguous in most cases that this is unlikely to have affected the overall pattern of findings.

The walkthrough, conducted on the Doubao web interface and informed by Light, Burgess and Duguay (2018), focused on design features mediating users' encounter with generated content: content moderation cues, reporting mechanisms, user response options, and platform-user communication channels. Observations were documented through screenshots and written notes.

Findings and Analysis

Spatial Coding and the Layered Othering of Regional Bodies

The context-free prompt set, [province] + people, revealed a stratified system of representation that goes well beyond simple economic mirroring. Across the 31 provinces, three distinct visual tiers emerged.

The first tier, comprising Beijing, Shanghai, Zhejiang, and Hubei, was characterised by modern or visually polished environments (urban skylines, scenic landmarks such as West Lake), a higher proportion of young adults than other tiers, and varied but well-fitted contemporary attire (Figure 1).

The national benchmark Chinese people (Figure 2) belongs alongside this tier in spirit if not in setting: although its scenes default to historical or heritage sites, the figures are predominantly young, dressed in carefully composed traditional or modern outfits, and aestheticised throughout. The first tier, in other words, is unified by polish, decency, and visual idealisation rather than by modernity alone.

A second, intermediate tier, including Guangdong, Jiangsu, Jiangxi, and Chongqing, was rendered through regionally specific cultural environments (Figure 3) : Cantonese morning tea settings, Jiangnan water towns, the Pavilion of Prince Teng, and Hongya Cave. These provinces avoided agricultural imagery, yet were rarely rendered through unambiguously modern urban environments either, occupying a middle representational space marked more by cultural distinctiveness than by either economic modernity or rural marginality.

The third and largest tier comprised the remaining 23 provinces (Figure 4) , where outputs converged on a strikingly consistent set of visual markers: agricultural settings, rural dwellings, ethnic minority costumes or worn cotton-padded jackets, predominantly middle-aged or elderly figures, and a recurring emphasis on weathered skin and lined faces.

Where the populations of the first tier were imagined as aestheticised, those of the second as culturally diverse, and those of the third as old, unkempt, and bound to the land—uniformly deviating from the idealised national benchmark. Therefore, AI's default image of "Chinese people" is not an inclusive composite of China's actual population. It is a narrow, idealised figure that excludes the majority of regional groups from its visual definition of Chineseness.

Beyond Economic Explanation: Labour and the Limits of GDP

If the context-free prompts revealed a stratified hierarchy of regional bodies, the [province] + people + at work prompts expose a deeper layer in which economic explanation alone proves insufficient.

Notably, eight provinces under this set—Gansu, Xinjiang, Tianjin, Shanxi, Chongqing, Shaanxi, Anhui, and Sichuan—were blocked outright by the platform at the point of generation (Figure 5) , returning no output and leaving these provinces outside the present analysis. Of the remaining 23, three distinct patterns emerged: seven provinces (Beijing, Shanghai, Hubei, Fujian, Jiangsu, Zhejiang, and Guangdong) generated almost exclusively white-collar imagery (Figure 6) , with figures in business attire operating computers in high-rise environments; six provinces (Guangxi, Guizhou, Yunnan, Liaoning, Shandong, and Jiangxi) produced mixed outputs in which manual labour predominated alongside occasional white-collar scenes; and the remaining provinces converged overwhelmingly on physical labour imagery: construction workers, factory operatives, and agricultural workers, accompanied by hard hats, workwear, and outdoor manual settings (Figure 8) .

Crucially, this distribution does not follow a linear relationship with GDP rank. The strongest evidence lies within the top ten provinces themselves, where representational outcomes split sharply. According to GDP rankings (ASKCI, 2025), Hubei (ranked 7th) and Fujian (ranked 8th) were rendered as white-collar workforces, while Henan, Hunan, and Shandong—of comparable economic standing—were depicted predominantly through manual labour (Figure 7) . Nor can this divergence be explained by industrial structure. Drawing on employment data from the China Statistical Yearbook (National Bureau of Statistics of China, 2024), I calculated that for Henan, Hunan, and Shandong, the tertiary sector accounts for a substantially higher share of employment than either agriculture or manufacturing. Yet Doubao's outputs almost entirely omit this structurally dominant service-sector workforce.

The case of Henan is particularly instructive. As Yin (2023) documents, Henan has long been subject to entrenched stigmatisation in Chinese internet culture, where its population is constructed as rural and uneducated. The persistence of manual labour imagery despite the province's diversified employment structure suggests that the model is not reflecting economic conditions but reproducing socially embedded perceptions. Peng (2021) demonstrates how Chinese news portals systematically amplify regional discrimination, deploying terms such as "fraudster" and "peasant" in association with provinces like Henan through affectively charged discourse. Since Doubao's training corpus is drawn from Chinese-language internet content, such media-circulated prejudices provide a plausible pathway through which historically sedimented stereotypes enter the model's representational repertoire. As Mehrabi et al. (2021) argue, biases embedded in large-scale datasets are not merely inherited but amplified through AI systems.

By contrast, the non-regional comparison prompts confirm the directional logic underlying these regional outputs. Chinese people at work consistently produced white-collar scenes, as did happy people at work and wealthy people at work; miserable people at work and poor people at work, by contrast, generated figures in workwear and manual settings. Doubao thus encodes a tight coupling between occupational category, emotional state, and class position: white-collar work is associated with happiness and wealth, physical labour with misery and poverty.

From Representational Harm to Digital Exclusion

Following Barocas et al.'s (2017) framework, the discrimination documented above is best understood as a representational harm: a form of injury that arises when certain groups are stigmatised or stereotyped, distinct from allocative harms in which resources or opportunities are directly withheld. Doubao does not bar users from stigmatised regions from accessing its services. The injury operates through use itself: when a user from one of the affected provinces recognises that the AI's depiction of their community as poor, rural, and occupationally inferior diverges sharply from lived reality and constitutes a form of discrimination, the experience may produce distrust, distress, and even psychological harm. Helsper (2012) identifies attitudes as a key mediator between social and digital exclusion: when users perceive a technology as not suited to their social group, disengagement is more likely. Building on this, sustained exposure to discriminatory representation erodes users' trust in the platform and, over time, drives them to reduce engagement with or abandon it altogether—producing a form of digital exclusion enacted not through technical barriers but through stigmatising representation itself.

An Architecture of Inaction

Public posts on Douyin began circulating evidence of Doubao's regional bias as early as February 2026, with users testing identical prompts across provinces and sharing the divergent outputs (Figure 9). Despite substantial engagement, no platform response followed, and the walkthrough conducted for this study suggests this persistence is not incidental.

Three observations are significant. First, no content warnings accompany generated images, leaving outputs to circulate as a neutral and natural result of the algorithm (Figure 10). Second, the reporting function is demoted: the report button is concealed within a collapsed three-dot menu rather than displayed alongside the image (Figure 11), raising the threshold of effort needed to register a complaint and reducing the likelihood that users will report problematic content. According to Doubao's Privacy Policy, such feedback is used either for general service improvement or—when the personalisation toggle is enabled—for refining personalised recommendations (Doubao, 2026). Neither use constitutes a dedicated channel for flagging systemic harms.

Taken together, these design choices constitute an architecture of inaction. The pattern of partial filtering reinforces this reading: eight provincial prompts were blocked outright at the point of generation, while the remaining provinces continued to yield clear stereotyping. Selectively silencing some publicly criticised cases while leaving the underlying generative logic untouched reflects a public-relations response rather than substantive accountability.

Conclusion

This study has documented systematic regional discrimination in Doubao's text-to-image function. The platform's outputs reproduce social prejudices rather than economic realities, most clearly in the divergent treatment of provinces with comparable GDP rank: Hubei and Fujian were rendered as white-collar workforces, while Henan, Hunan, and Shandong—of similar economic standing—were depicted through manual labour. The pattern operates on two levels. At the level of the model, it appears as systematic bias in data, algorithmic outputs, and visual representations. At the level of governance, it appears as a platform that retains the capacity to intervene—as the selective filtering of eight provincial prompts demonstrates—yet declines to address the underlying logic that produces discriminatory outputs.

These representational harms produce digital exclusion through their effect on users' trust: when users recognise that a platform encodes their community as poor, rural, and inferior, sustained engagement becomes harder to maintain. As an exploratory, single-coder study, this research identifies patterns rather than quantifying them. Nevertheless, by integrating literatures on algorithmic bias, China's regional inequality, and digital exclusion, this paper documents a culturally specific form of AI-generated harm. As generative AI becomes ever more embedded in everyday life, the question of whose experiences these systems encode—and whose they marginalise—cannot be left to platforms alone.