Exploring Cultural Bias in AI-Generated Images of East Asian Women: Methodology, Findings, and Implications

Study Overview and Objectives

Generative artificial intelligence (AI)—particularly image-generating models—has come to play a significant role in shaping visual representations across digital media, creative industries, and public discourse. Yet, these technologies do not merely reflect objective reality; instead, they risk amplifying entrenched cultural, racial, and gender biases already present in society. Notably, while AI bias has received considerable scrutiny in Western contexts, systematic and nuanced analyses of how such bias impacts the portrayal of East Asian women remain scarce. The study authored by Xingyu Lan, Jiaxi An, Yisu Guo, Chiyou Tong, Xintong Cai, and colleagues breaks ground in this domain by offering a comprehensive audit of three popular image-generating AI models—OpenAI’s DALL-E 3, Midjourney v6, and Stability AI’s Stable Diffusion 3—with a focus on the depiction of East Asian women.

The primary aim of the research is to systematically surface, classify, and analyze the perceived cultural, racial, and gender biases embedded in AI-generated visuals of East Asian women. Through this lens, the study seeks both to highlight the representational challenges to identity, dignity, and cultural authenticity and to provide actionable directions for developing more responsible AI systems.

Key objectives guiding the research include:

Auditing leading AI image-generation models for cultural biases in depictions of East Asian women
Establishing a taxonomy of commonly occurring biases and their visual manifestations
Comparing perceived biases across different AI platforms and user demographics
Investigating broader societal, technological, and ethical ramifications
Offering grounded recommendations for policy, industry, and future research

This exploratory work not only addresses significant gaps in the literature but also responds to real-world controversies involving cultural misrepresentation and offense, as seen in recent years with major brands and viral social media content.

Methodology of the Study

Algorithmic User-Centered Audit

To ensure authentic and contextually relevant assessment of biases, the study adopted a user-driven algorithm audit method. This approach emphasizes evaluation by those most impacted: real-world, self-identified East Asian users. Auditors engaged directly with DALL-E 3 (OpenAI), Midjourney v6, and Stable Diffusion 3 (Stability AI), which together dominate the text-to-image generative space and have broad public availability and influence.

Participant Recruitment and Demographics

Participants were recruited via social media outreach and snowball sampling strategies with eligibility criteria that ensured both demographic relevance and digital literacy:

Self-identification as East Asian (Chinese, Japanese, Korean)
Current residence within East Asia
English proficiency (to ensure engagement with international AI tools)
Prior experience using image-generating AI

In total, 36 participants took part (25 females and 11 males), spanning ages 18–50 (mean: 23, SD: 6.3). The group included 20 Chinese, 7 Japanese, and 9 Koreans. All participants held at least a college degree, with academic backgrounds spread across social sciences, engineering, medicine, and management.

Study Procedure and Data Collection

Each participant was introduced to the study’s background and underwent brief tutorials in using the three AI tools. They were then asked to generate images using prompts centered around “attractive East Asian woman,” with flexibility to customize adjectives and cultural referents.

The process included:

Generating images with all three AI models
Recording prompts, tool types, and outcomes
Providing “ideal” example images reflecting their own cultural preferences
Reviewing and annotating perceived biases in AI-generated results
Participating in semi-structured interviews on beauty ideals, satisfaction, and discomfort
Quantitatively rating model satisfaction on a 0–100 scale

Qualitative feedback was systematically coded to identify and aggregate recurring themes, while quantitative ratings allowed for model comparison.

Audit Focus

The audit focused specifically on depictions of women, given public salience and heightened controversy surrounding female representation in both AI-generated and traditional media. Male imagery was not directly assessed in this phase, in acknowledgment that distinct interpretive histories and representational tropes shape perceptions of East Asian men.

Participant Recruitment and Demographics (Expanded)

The demographic composition of the study audience is critically important, as it ensures that culturally situated perspectives inform the interpretation of model outputs. Recruitment took pains to include East Asians living within their home cultures, counterbalancing the Western-centric bias commonly found in AI research and audit literature.

Sample characteristics:

Country representation: Chinese (20), Japanese (7), Korean (9)
Gender: 69% female, 31% male (reflecting heightened public attention to the subjectivity and objectification of women in media and technology)
Age group: 18–50; average 23, reflecting the digitally fluent, social media-engaged demographic most likely to use and interact with generative AI
Educational background: All college-educated, enabling nuanced discussion of both technical and cultural aspects

All participants self-identified as culturally “insider” and expressed confidence in evaluating cultural accuracy and nuance. Their input thus offers a counterweight to Western subjectivities that often dominate the shaping and critique of global AI models.

Key Findings

Ideals of East Asian Female Beauty

Participants articulated a distinct set of indigenous beauty standards: emphasis on facial features (notably, luminous, harmoniously proportioned eyes), well-balanced facial shape (generally round or oval rather than angular), fair but natural skin, minimal makeup, and a soft, dignified temperament. Traits like youthfulness were appreciated in moderation, while strong sexual or hyper-feminine signals were not universally admired. Elegance and naturalness—rather than Western-defined glamour or overt femininity—emerged as key ideals.

Despite this, only 43.7% of AI-generated images (across all models) were rated by participants as even moderately “satisfactory.” This result signals substantial cultural misalignment in the creative logic of these platforms.

Model satisfaction mean scores:

Midjourney: 72.2
DALL-E: 62.8
Stable Diffusion: 52.3

Taxonomy of Perceived Biases

Eighteen distinct types of perceived bias were identified and systematically grouped into four key patterns:

Bias Pattern	Issues/Examples	Core Manifestations	Leading Model Source
Westernization	Makeup, sharp face, deep eyes, high/narrow nose, full lips, tanned skin, big smile	Overrepresentation of Eurocentric features and beauty standards, heavy or Western-style makeup, angular face, pronounced nose bridge, exaggerated toothy smile, bronzed rather than natural skin tone, deeply set eyes	DALL-E
Sexualization & Feminization	Revealing clothing, porcelain skin, youth, dull/harsh gaze, slenderness, long hair	Depiction of women as youthful, thin, docile, submissive; sexualized attire (deep necklines, short skirts), overidealized skin texture, lifeless or overtly harsh gazes, passive or seductive posture	Stable Diffusion
Overuse/Misuse of Cultural Symbols	Ancient/exotic symbols, mixed cultural elements, incorrect context	Indiscriminate or incorrect use of “traditional” Asian motifs (e.g., lanterns, ornate attire, temples), confusion between Chinese/Japanese/Korean signifiers, use of props or symbols out of context	Midjourney
Racial Stereotype	Slanted/narrow, upturned eyes	Reinforcement of historic racial caricatures; eyes depicted as sharply upturned or unnaturally narrow, evoking offensive tropes perpetuated in Western art, media, and advertising	Stable Diffusion, DALL-E

Table elaboration: The Westernization pattern dominated the overall distribution of reported biases (144 coded items), followed by sexualization and feminization (64), misuse of cultural symbols (63), and finally racial stereotype (13).

Model-level distinctions highlight that DALL-E outputs were most strongly associated with Westernized forms of bias; Midjourney was specifically prone to cultural confusion and symbol misuse; Stable Diffusion frequently generated sexualized or racially stereotyped images (see also independent comparative benchmarks).

Table: Key Biases Identified and Their Manifestations in AI-Generated Images

Bias Pattern	Specific Issues	Description / Manifestation
Westernization	Western makeup, sharp/bony face, deep eyes, high/narrow nose, tanned or lightened skin, wide smile	Images adopt Western beauty standards: heavy makeup, angular faces, pronounced cheekbones/nose, skin that is either paler or artificially tanned, exaggerated smiles
Overuse/Misuse of Symbols	Exotic/ancient artifacts, mixed elements, wrong semantics	Traditions, props, attire appear out of context; incorrect blending of elements from different Asian cultures; symbols are used as aesthetic ornaments without meaning
Sexualization/Feminization	Revealing attire, porcelain skin, youth, passive or seductive expressions	Images depict women as thin, young, docile; overemphasis on submissiveness; clothing reveals more skin than is culturally normative; lifeless or hyper-feminine expressions
Racial Stereotype	Slanted eyes, exaggerated ethnic features	Upturned, narrowed eye shapes align with colonialist caricatures; reinforce “orientalist” imagery; skin tones and features often homogenized or “whitewashed”
Representation Stereotypes	Uniformity, lack of individuality, generic attire	East Asian women are visualized as a monolithic type, with little attention to intra-regional diversity in appearance, dress, or setting
Contextual Erasure	Unrepresentative group/mixed-race depictions	In multi-person images, East Asian women are erased, lightened, or replaced by white-coded features
White Normativity	Overrepresentation of white features or aesthetics	Skew in image outputs toward Caucasian or “global North” standards of attractiveness and presentation

(References: )

Comparative Analysis Across Models

Direct model comparison revealed distinct patterns of bias and model behavior:

DALL-E: Most Westernization bias; frequently generated women with heavy Western-style makeup, angular features, and broad smiles.
Midjourney: Most frequent misuse and overuse of traditional cultural symbols and blending of motifs from different East Asian regions; sometimes blurred national boundaries or placed inauthentic, random symbols in modern or unrealistic contexts.
Stable Diffusion: Showed strongest tendency toward sexualization, hyper-feminization, and racial stereotyping; outputs included more overtly revealing clothing, thin/stereotypical body shapes, and frequent use of slanted/upturned eye motifs.

Participant satisfaction reflected these trends—with users expressing marked discomfort when images strayed furthest from indigenous aesthetic ideals, or when images replicated historic offensive tropes.

Broader Societal Implications

Cultural Alienation and Discomfort

The pervasiveness of Westernized or sexualized depictions of East Asian women has substantial psychosocial effects. For East Asian women, these images can elicit emotions of cultural alienation or discomfort; when one’s identity is persistently filtered through a Western gaze or commodified as an “exotic other,” it undermines personal autonomy, pride, and dignity. These AI-generated images serve as new vectors for long-standing global power imbalances, with Western media and technology industries reinforcing mono-cultural beauty ideals—even within tools intended for global or diverse audiences.

Youth and young adults, particularly those navigating identity in cosmopolitan, internationalized settings, are uniquely vulnerable to these forms of digital misrepresentation. The effects of “white-washing” and colorism—where fair skin and Eurocentric features are privileged over genuine ethnic diversity—have been well documented across social media trends and online communities.

Reinforcement of Stereotypes and Historical Harm

AI-generated images that perpetuate tropes such as the “China Doll,” “Dragon Lady,” or hypersexualized “Geisha Girl” contribute to a cycle of cultural and sexual objectification with roots in colonial and orientalist imagery. As numerous scholars and activists have warned, media and algorithmic repetition of such stereotypes not only distorts public perception but also legitimizes discrimination and violence against East Asian women—manifested in both microaggressions and extreme hate crimes (e.g., Atlanta spa shootings, widespread reports of harassment during and after the COVID-19 pandemic).

Cultural Appropriation and Misuse

Uncritical or insensitive recombination of overt cultural signifiers (e.g., blending Chinese, Japanese, and Korean dress and props) results in the flattening of rich and diverse traditions. Rather than celebrating authentic difference, AI-generated “fusion” risks cultural appropriation, reinforcing orientalist tropes, and even erasing intra-Asian differences. This simplification upholds “model minority” or “exotic fetish” narratives rather than allowing space for self-determinate cultural expression.

Feedback Loops and Digital Inequality

When AI-generated representations become part of new training sets or shared widely on social media, they may further entrench new rounds of bias: biased output becomes biased training data, fueling a vicious cycle. As AI-generated images increasingly populate search results, marketing materials, and even news reporting, they risk becoming new “benchmarks” for ethnic or gender identity, which can have pernicious effects on social equality and inclusion practices.

Technological and Ethical Implications

Data Skew and Training Set Bias

A key technological driver of these biases is the skewed nature of AI training datasets. Most models (e.g., LAION-5B) are dominated by images and captions from the “global North”—primarily Western, English-language sources. For example, a cited dataset had 18 times more Western than non-Western references, with 78% of image-text pairs coming from North America/Europe and less than 7% from Asia or Africa. As a result, the “neutral” or “default” output of these systems privileges Western aesthetics and patterns, even when representing non-Western subjects.

Tokenization and model architecture contribute further limits, with semantic encoding schemes often unable to distinguish subtle cultural and intra-regional differences. Algorithms may collapse or blend concepts for efficiency, resulting in the loss of meaningful cultural nuance.

Algorithmic Blindness and Evaluation Shortcomings

Mainstream AI evaluation metrics (like FID or CLIP Score) privilege photorealism or semantic alignment but fail to capture cultural fidelity or context-specific fairness. These metrics may thus “approve” images that are photorealistically rendered yet perpetuate stereotypes or inaccuracies. Recent attempts, such as the Component Inclusion Score (CIS), seek to quantify cultural appropriateness, but even these can import bias depending on the reference corpus.

Model “guardrails” and content filters are mostly designed to catch explicit violence or sexual material and rarely address subtler cultural or representational harms. Consequently, Westernization, cultural flattening, and aesthetic bias persist even as more overt (NSFW) bias may be filtered out.

Cross-Model Disparities

Open-source models (e.g., Stable Diffusion) and closed-source models (e.g., DALL-E, Midjourney) differ in their mitigation approaches and content policies. DALL-E has invested in ethical review, content moderation, and provenance tools, resulting in more frequent filtering or rejection of overtly problematic prompts. However, this filtering can also erase complex representation and “overcorrect,” producing images that fall into a different form of bias—white-washing or extreme inoffensiveness. Midjourney and similar platforms, prioritizing artistic freedom and visual diversity, often allow a wide spectrum of output—sometimes resulting in richer results, but also permitting unchecked bias and cultural confusion.

Ethical Accountability and Transparency

The rapid global diffusion of generative AI has far outpaced the development of robust ethical review and transparency practices. Companies’ reluctance to disclose training data composition and filtering mechanisms—and the proprietary nature of many foundational models—make it difficult to fully audit and mitigate embedded bias. In practice, industry self-regulation has proved insufficient for protecting marginalized populations or safeguarding cultural authenticity.

Recommendations for Mitigation

Table: Recommendations for Bias Mitigation in AI Image Generation

Recommendation Category	Proposed Action
Data Diversity and Quality	Proactively curate regionally diverse visual datasets, including authentic modern and traditional images of East Asian women promulgated by local cultural organizations, artists, and communities. Fund and support data-collection initiatives in underrepresented regions.
Algorithmic Oversight	Audit models with input from affected communities. Develop model architectures or plugin tools allowing user control over aesthetic and cultural parameters. Implement periodic algorithmic bias testing and “red teaming” to stress-test cultural boundaries.
Human-in-the-Loop Evaluation	Recruit culturally diverse human evaluators and employ participatory design for prompt testing, filter review, and content flagging. Engage East Asian artists, social scientists, and community leaders directly in the audit process.
Transparency Measures	Disclose both training data sources and known data composition flaws. Label AI-generated imagery visibly, with context about model and data origin; provide users with explanations on why certain content has been filtered or flagged.
Cultural Sensitivity Filters	Develop and integrate prompts, filters, and post-processing tools to catch mistaken or offensive symbol usage, prompt content checks for historicized stereotypes, and allow users to select or veto certain motifs (e.g., “no cultural fusion” option).
Regulatory and Policy Reform	Support national and international regulation mandating AI model audits for racial/gender/cultural bias, labeling requirements for AI-generated content, and opt-out provisions for data subjects. Reference and build upon frameworks like the UNESCO AI Ethics Recommendations (2021), EU AI Act, and regional regulatory initiatives in Asia.
Education and Public Literacy	Expand AI literacy programs for the public and specialist communities. Educate both developers and creators on the risks of bias, cultural appropriation, and stereotype propagation, emphasizing the human impacts of algorithmic harm.

Recommendations Explained

Diversify Training Data
Ethical AI model development requires the inclusion of representative datasets built in partnership with local communities, artists, and cultural stakeholders. Datasets such as LAION-5B and others must be expanded and regularly updated to counteract Western-centric bias and fill gaps in the visual record for underrepresented groups.
Algorithmic Auditing, Transparency, and Reporting
Regular, multidisciplinary algorithm audits should become standard. These audits should track model behavior across race, gender, and cultural lines, using context-aware benchmarks. Reporting mechanisms and transparency protocols—similar to food labeling or financial audits—should be required for commercial platforms and open-source models alike.
Community and Artist Involvement
Involving East Asian artists, designers, activists, and scholars from the target communities in dataset curation, model design, and bias evaluation will help ensure that representation is rooted in lived experience and not imposed from a global North perspective.
Model-Embedded Cultural Controls and User Feedback Loops
Models should incorporate region and culture-specific parameterization, with user-controllable options for specifying desired or undesired motifs. Prompts and feedback channels should enable the flagging, reporting, or correction of inappropriate, offensive, or misleading outputs in real time.
Policy Intervention and Regulatory Alignment
Governments must establish requirements for fairness, representational accuracy, and cultural transparency in both public sector and commercial AI deployments. International frameworks—such as the UNESCO Convention on the Protection and Promotion of the Diversity of Cultural Expressions—offer a foundation for policy-setting. For high-impact use cases, regulatory “red lines” could ban the commercial use of models that fail minimum diversity and representation benchmarks.

Policy and Regulation

Regulatory Landscape

National and international organizations have begun to recognize the threat of AI-driven divisiveness, erasure, and discrimination. The EU AI Act (2024), for example, mandates risk-based classification and bias audits for high-impact applications. China and Singapore have taken steps by requiring algorithm traceability, bias self-assessment, and (in Singapore) “AI Verify” toolkits for model auditing and transparency.

UNESCO’s Recommendation on the Ethics of Artificial Intelligence (2021) establishes cultural diversity and human-centered design as guiding pillars. It calls for:

Consent for data contribution
Fair remuneration and intellectual property protection
Transparency in training data, prompt design, and output
Development and public disclosure of inclusive, regionally representative benchmarks.

Key Recommendations

Labelling of all AI-generated outputs should become mandatory for visual media, with provenance and data lineage specified.
Global protocols for dataset opt-out and content reporting must be implemented.
Policy harmonization across regions is needed to prevent regulatory arbitrage and data colonization.
Governmental support should be extended toward public, “high-quality” dataset development sourced ethically from minority or underrepresented populations, with financial mechanisms (e.g., creative industry taxes) supporting local data enterprises.
Cross-sectoral educational programs should build AI literacy and diversify the developer pipeline.

Regulators are also urged to consider the intersectionality of harm—ensuring that audits account for the overlapping effects of race, gender, age, class, and other lines of marginalization.

Conclusions

The work of Xingyu Lan, Jiaxi An, Yisu Guo, Chiyou Tong, Xintong Cai, and their collaborators represents a crucial step in exposing the hidden mechanics by which contemporary generative AI preserves, amplifies, and even invents new forms of cultural, racial, and gender bias. Their findings reveal not simply a technical failure, but a wider societal and ethical challenge—one that requires urgent, multidisciplinary intervention.

AI-generated images of East Asian women reflect and reinforce dominant Western beauty standards, sexualization patterns, cultural confusion, and racial stereotypes. These outputs arise from both technical artifacts (data imbalance and architectural constraints) and sociotechnical systems (commercial priorities, global digital colonialism, weak regulatory frameworks). They contribute to real-world harms: alienation, diminished cultural identity, feedback loops of stereotype propagation, and the erasure of authentic voices and histories.

Mitigating these harms demands a multilayered response: dataset reform, algorithm auditing, participatory design, transparency, regulatory enforcement, and ongoing public engagement and education.

If neglected, AI risks entrenching the very hierarchies global technology has the power to disrupt. If addressed with rigor, empathy, and collaboration, generative AI could instead become a tool for wider cultural self-determination and global equity.