Systemic Bias in Occupational Gender Representations in China: A Cross-Platform Audit of Search Engines and Generative AI

Abstract

As AI permeates daily life, algorithmic platforms increasingly function as complex sociotechnical systems that shape public perception and societal attitudes. Addressing concerns that AI text-to-image models and search engines reinforce stereotypes, this study focuses on China, a context marked by traditional gender norms and a vast technological ecosystem, examining how algorithmic systems perpetuate gender power structures through occupational representations. Using algorithmic audits of 60 occupations, Z-tests, and QAP network analysis, this study compares platform gender representations with national census data, systematically distinguishing “generative bias” in AI platforms (Doubao Seedream 3.0, Jimeng Image 3.0) from “retrieval bias” in search engines (Baidu, Sogou). Findings reveal that search engines reinforce stereotypes by over-representing dominant genders and obscuring non-mainstream ones. Generative AI exhibits more radical distortions. The specialized AI Jimeng shows a strong gender polarization feature, while the general AI Doubao shows an ideal balanced gender presentation tendency, balancing representation yet creating an equally false reality. Compared to search engines, AI platforms have greater creativity in representing occupational gender. This study reveals a mutually reinforcing bias cycle among audiences, media, and algorithms, offering a crucial non-Western perspective for feminist technology studies and significant implications for equitable AI governance.

1. Introduction

As artificial intelligence systems become increasingly embedded in everyday social life, platform algorithms have become pivotal forces shaping public cognition and societal perceptions. From a systems perspective, these algorithmic platforms do not operate in isolation but function as interconnected nodes within broader sociotechnical ecosystems, where biases propagate through feedback loops across multiple platform types. NVIDIA’s withdrawal from China’s high-end chip market in October 2025 further underscores major economies’ accelerating drive toward technological self-reliance in critical domains. Against this backdrop, China’s market scale and distinctiveness offer a crucial arena for algorithmic research. By June 2025, China’s generative AI user base had exceeded 515 million, achieving a 106.6% growth rate over six months. According to a survey released by China Internet Network Information Center (CNNIC), over 90% of users will first opt for large language models developed domestically [1].

However, with the advancement of artificial intelligence, the systemic issue of fairness has also drawn widespread attention. In the interplay among users, data, and algorithms, stereotypes may be incorporated into algorithms through training data and user behavior, and subsequently re-presented to users via these algorithms [2]. In AI systems, bias typically manifests as systemic errors, leading to unequal treatment of different individuals or groups. Such bias includes statistical biases (such as selection bias, measurement bias, and algorithmic bias) as well as social biases that reflect historical inequalities, social stereotypes, and differences in evaluation [3]. As visual representations of social reality, images often exhibit stronger communicative efficacy and emotional resonance than textual forms. Their inherent biases also influence audience perceptions of gender and occupation more directly than written language [4]. Today, AI text-to-image technology is rapidly integrating into mainstream visual culture production systems, emerging as a vital medium for digital communication [5]. In the Chinese market, general-purpose AI platforms like Doubao, Tencent Hunyuan, and Tongyi Qianwen have successively integrated text-to-image capabilities, while specialized platforms such as Jimeng and Keling have also emerged. This vast and highly endogenous technological ecosystem provides ample empirical conditions for examining the cross-cultural manifestations and impacts of algorithmic bias.

In fact, algorithmic bias is not unique to generative AI. Search engine image results have also been criticized for reinforcing gender stereotypes through ranking and recommendation mechanisms. Studies on Western platforms like DALL-E and Google have confirmed their tendency to perpetuate and amplify societal biases, particularly regarding to occupational gender roles [6,7,8]. From a sociological perspective, such technological bias is closely linked to the theory of “occupational gender segregation.” Historically, women have been concentrated in “emotional-labor” roles such as caregiving, education, and administrative work, while men have dominated in “power-labor” fields like technology, engineering and management [9]. Algorithms often unconsciously learn and reproduce this data structure, leading to the “technological” reproduction of occupational gender segregation. However, current research perspectives are predominantly grounded in Western platforms and data, requiring verification of their conclusions’ universality across diverse technological environments and sociocultural contexts. In Asian regions like China, traditional gender norms such as “male supremacy and female subordination” and “three obediences and four virtues” remain deeply entrenched. Confucian frameworks have historically emphasized a gendered division of labor, assigning men to public and external affairs and women to domestic and internal roles [10]. As a result, women often encounter persistent normative constraints when navigating public evaluations of their professional performance [11,12], manifested in stereotypes, role conflicts, and limited social support [13]. Whether these culturally embedded norms give rise to distinct patterns of algorithmic bias in occupational gender representation, as compared to Western contexts, warrants systematic empirical investigation.

Based on this, this study employs algorithmic auditing methods, focusing on China’s mainstream AI platforms and search engines, to propose two new analytical dimensions: “generative bias” and “retrieval bias”. We use the term “content biases”, as mentioned in existing research, to refer to systematic deviations in retrieved information relative to known or accepted benchmarks of truth. Content bias may stem from three sources: (i) creators of online content; (ii) search engines, which may index existing content and retrieve a biased set of results; (iii) searchers, who are influenced by their own biases when searching, perceiving, and reacting to the presented information [14]. This content bias negatively impacts the accuracy of results. Although previous research on content bias has primarily focused on the fields of web search and information retrieval, in this study we extend this concept to text-to-image AI models, defining generative bias as the deviation of AI-generated images from reality. As awareness of systemic flaws in AI technology deepens, algorithmic bias has garnered significant attention as one such flaw. It occurs “when an algorithm’s output benefits certain individuals or groups or disadvantages others without justifiable reason” [15]. AI algorithms may develop “generative content bias” through the training data and sampling processes used to synthesize outputs. Therefore, in this study, “generative content bias” (hereinafter collectively referred to as “generative bias”) refers to the occupational and gender representation tendencies exhibited by generative AI during the image generation process. In contrast, search engines primarily exhibit retrieval-related content bias through processes such as indexing, ranking, and query interpretation. Therefore, this paper collectively refers to biases occurring in the indexing, ranking, and recommendation stages that influence the presented results as “retrieval content bias” (hereinafter collectively referred to as “retrieval bias”). In this study, retrieval bias specifically refers to the deviation between the occupational gender images presented by search engine platforms and the actual occupational gender structure. Therefore, generative bias reflects the “creating from scratch” aspect of content production, while retrieval bias reflects the “highlighting existing content” aspect of content distribution. These two concepts correspond to two distinct types of platform technologies and illustrate the differing pathways through which the final image is presented.

Taking occupational gender bias as its primary analytical focus, the research examines gender ratios in occupational images generated or retrieved by algorithmic platforms to explore the differences in bias manifestations and mechanisms between the two types of algorithms. The paper aims to address the following questions:

RQ1: Do AI platforms and search engines exhibit generative bias and retrieval bias, respectively? How do these biases manifest, and to what extent do they converge or diverge?

RQ2: Compared to real-world data, which type of bias (generative bias in AI platforms or retrieval bias in search engines) is more likely to amplify existing bias?

RQ3: What sociocultural norms and gendered occupational meanings are reflected through these platform-based visual representations?

Through empirical analysis of China’s digital ecosystem, this study aims to provide a much-needed Eastern perspective for understanding the cross-cultural manifestations and impacts of algorithmic bias. It advances a cyclical feedback system for understanding how algorithmic bias emerges from, and feeds back into, the interplay among technological design, sociocultural norms, and human cognition within the information ecosystem. It contributes critical evidence toward building an inclusive and equitable global AI governance paradigm, thereby advancing feminist scholarship on algorithmic power in the digital age.

2. Literature Review

2.1. Occupational Gender Structure and Technological System Bias

Occupational gender structure refers to the uneven distribution of men and women across occupations in the labor market. This structure is not naturally formed but rather constructed through social cultural coding and value stratification processes, interacting reciprocally with gender stereotypes in social cognition [16]. Gender stereotypes, as a sociopsychological phenomenon, represent stable expectations and generalized perceptions people hold about the behaviors and personality traits of men and women. Once established, these stereotypes exert broad influence on individuals’ psychology and behavior. With advances in media technology, gender stereotypes are no longer maintained solely by families and educational systems. These ideas are reinforced implicitly through media narratives, films, television programs, and digital communication processes [17].

In traditional media environments, portrayals of occupational gender roles in advertising, news, and film/television serve as key mechanisms reinforcing the notion that “men are the economic providers while women are homemakers” [18]. In the era of smart communication, this “occupational gender coding” persists. Social media, search engines, and artificial intelligence have been defined by scholars as new vessels for stereotypes [19,20], reinforcing specific gender images through data input, algorithmic recommendations, and content filtering—subtly embedding gender bias into the technical ecosystem [21,22]. This process transforms stereotypes into technical system biases. Technology is not value-neutral. Social values and inherent biases interact with technical elements such as algorithms, and prompts through various groups including developers and users, and are deeply integrated into the technical system [23]. However, contrary to public expectations, research reveals that people generally perceive technologies like algorithms as more fair than humans, often responding more leniently to biased outcomes generated by technology than to human-based bias [24]. Therefore, exploratory empirical studies on technological bias are urgently needed to uncover discriminatory issues and effectively prevent the negative impacts brought by technology.

2.2. Mechanisms and Manifestations of Algorithmic Bias

As society’s reliance on technology and artificial intelligence deepens, various related issues have drawn increasing attention, among them algorithmic bias. Algorithmic bias refers to discriminatory behaviors toward different groups caused by computer algorithms [25]. Existing research identifies three stages in the emergence of algorithmic bias: the training process using input data, the platform’s algorithm development process, and the actual application of algorithms [26]. Therefore, empirical studies focusing on the biased outcomes directly affecting audiences represent one effective approach to understanding algorithmic bias. The practical application of algorithms is most visibly reflected in the construction of platform visibility—platforms use recommendation systems, ranking logic, and generative models to determine “who gets seen,” thereby intervening in the ongoing formation of gender norms.

Studies have shown that search engines and artificial intelligence models, which are the two most widely used types of algorithms today, are exacerbating age-related gender bias [27]. This study focuses specifically on gender bias in these algorithms as it relates to occupational roles. Search engines are a classic example of “retrieval bias,” where search algorithms sort and recommend content in ways that produce biased information displays, reinforcing users’ pre-existing cognitive patterns regarding gender, race, and politics [28]. Research reveals systematic bias in Google Images’ portrayal of gender distribution across professions. For instance, women account for only 11% of images associated with high-status positions such as chief executive officer, despite making up approximately 27% of CEOs in the United States. Similar gender bias has also been observed in China’s Baidu search engine [29]. Beyond search engines, AI models exhibit what scholars describe as “generative bias.” For example, women are more frequently predicted to occupy roles such as “home care,” while men are more often associated with senior management and technical positions [30]. Both “retrieval bias” and “generative bias” lead to certain genders being overrepresented or marginalized in specific occupations. These biases may then feed back into the data ecosystem through generated content, reinforcing a “cognitive loop” that perpetuates occupational gender stereotypes among audiences.

2.3. Occupational Gender Representation in Digital Visual Systems

Algorithms are a set of mathematical instructions designed to achieve specific goals. All kinds of visual information in images need to be transformed into a vector-based digital framework before they can be recognized and interpreted by machines [31]. The binary gender concept is simple and intuitive, and it can be easily converted into data and incorporated into the model, thereby giving rise to gender bias at the algorithmic level. Studies have shown that humans are far more efficient in receiving and processing multimodal information such as images than in processing pure text content. This characteristic makes gender-biased visual content more likely to be deeply ingrained in people’s minds, ultimately leading to more severe gender bias in images than in text [27].

The stereotypes in technological bias gradually evolve into new hidden social forces, causing systematic gender discrimination for women in the digital environment [32]. A large number of empirical studies have shown that visual gender stereotypes can weaken women’s career aspirations and self-efficacy, especially in senior management and technical positions [33]. Women in the workplace will encounter the “glass ceiling” phenomenon and face numerous obstacles in promotion and leadership positions. Scholars have pointed out that such social biases and technology mutually shape each other [34]. Therefore, it is necessary to expand the feminist perspective to the understanding of the essence of technology. Feminist technological theory holds that technology is not a neutral tool but is embedded with social gender power relations during the design, production, and application processes. Exploring occupational gender biases from a feminist technological perspective is crucial for achieving gender equality [35]. Some scholars have proposed technical solutions, including training data review and balanced sampling, reducing the prompts for stereotypical outputs, and post-processing algorithms for content review and reducing bias [36,37]. However, no single measure can eliminate algorithmic bias. Therefore, interdisciplinary collaboration is crucial for reducing gender bias in visual representations [38].

In summary, the current research mainly focuses on how generative artificial intelligence and search engine algorithms perpetuate and strengthen occupational stereotypes, particularly in terms of occupational image, age, gender, and occupational distribution, among other biases. However, the related research still has several limitations. Firstly, most existing studies mainly explore the generative biases of artificial intelligence models and the retrieval biases of search engines separately, with few integrating the two for comparative analysis to comprehensively reveal the different dimensions of algorithmic biases [39,40,41]. In addition, current research on biases in artificial intelligence and search engines mainly focuses on the international (especially Western) context, while research from Asian regions such as China is relatively limited.

In this context, this paper employs algorithmic auditing methods to conduct research on China’s AI text-to-image platforms and search engines, systematically analyzing and comparing the occupational gender biases present in these platforms. This study reveals how algorithmic technologies interact with specific social and cultural factors within the Eastern cultural background, resulting in a bias pattern that is completely different from those observed in Western research. Additionally, from a social science perspective and a feminist technology perspective, the practical application results of algorithms are discussed, which is helpful in systematically understanding algorithmic biases from a cross-disciplinary perspective outside of computer science. This can improve the exploration of algorithmic fairness in various disciplines and prompt people to pay attention to the power relations and capital factors behind technological systems, promoting the formation of a responsible, open, and inclusive technological ecosystem [42,43].

3. Methods

3.1. Research Subjects and Sample Selection

According to the “2025 China Mobile Internet AIGC Sector Traffic Report” released by iResearch, Doubao (The international title is Dola and the original title was Cici) has become a leading large language model in China, and Jimeng (The international title is Dreamina), a professional image platform, has occupied 95% of the image processing market [44]. Regarding Chinese search engines, a comprehensive analysis of mainstream information sources is conducted by choosing Baidu (the leader in the Chinese market) and Sogou (a representative Chinese search engine), and account for over 65% of the market share of Chinese search engines [45]. Most previous studies have been case studies of specific search engines or artificial intelligence tools. For instance, some scholars conducted research on the Midjourney single platform, indicating that it tends to generate images featuring mainly men [40]. There were also studies conducted on the four major search engines, namely Google, Bing, Baidu and Yandex, revealing that all four platforms have image biases that are unfavorable to women [39]. Therefore, selecting these four representative platforms—Doubao, Jimeng, Baidu, and Sogu is a practical and feasible approach for this study. We selected the Chinese-localized versions of these four platforms as the subjects of our study. Among them, the two AI platforms are Doubao Seedream 3.0 (hereinafter referred to as “Doubao”) and Jimeng Image 3.0 (hereinafter referred to as “Jimeng”).

Due to the lack of published data from the Seventh National Population Census and industry-specific statistics, comprehensive gender distribution figures at the occupational subcategory level are unavailable. Therefore, this study adheres to the methodological convention in China’s algorithmic auditing research—the occupational categories and gender distribution data required for the study were sourced from the “Nationally Employed Population by Gender, Occupational Subcategory, and Weekly Working Hours” statistics in China’s Sixth National Population Census [2,46]. The occupational classifications in this dataset adhere to those specified in the Occupational Classification Code of the People’s Republic of China. Given that the names of this occupational category are predominantly formal in tone, this study centers on official standard occupational titles, while employing commonly used colloquial terms for certain complex titles to facilitate retrieval or generation. For instance, the input term “wood product producers” was modified to “carpenters” to align with multi-channel data representation practices and enhance image acquisition accuracy. The census dataset comprises 409 occupational subcategories, along with their respective gender ratios. Subsequently, occupational subcategories with overly granular classifications or highly similar visual characteristics were excluded. Occupations with insufficient search engine image results or suboptimal AI generation outcomes were also eliminated, ultimately retaining 184 occupational subcategories. These occupations were ranked in descending order according to the proportion of female employees. To capture occupations with distinct gender distributions, we selected three groups of occupations based on their ranking positions: the top 20 occupations with the highest female representation, the bottom 20 occupations with the lowest female representation, and 20 occupations located around the median range of the ranking. As shown in

Table 1. 60 Occupations and Their Female Proportions (%).

High Female Proportion	Preschool Teacher	Nurse	Childcare Provider	Accountant	Cleaner
	93.3	93.1	89.2	73.1	71.6
	Seamstress	Librarian & Archivist	Hospitality Server	Statistician	Translator
	71.0	69.8	67.9	67.3	66.8
	Announcer/Program Host	Aquatic Product Processor	Beauty/Hairdresser	Nursing Aide	Pharmacist
	61.4	61.2	60.6	60.5	60.4
	Primary Teacher	Special Education Teacher	Toy Maker	Inspector	Tourism/Recreation Attendant
	59.7	59.7	59.5	58.8	58.4
Medium Female Proportion	Journalist	Relics Protector	TCM Physician	Printing Operator	Rubber Producer
	41.1	40.7	40.6	40.0	38.0
	Sociologist	Securities Professional	Plant Protector	Environmental Monitor	Crafts Professional
	38.0	37.3	37.1	37.1	36.6
	Funeral Attendant	Agricultural Science Researcher	Film/TV/Stage Producer	Property Manager	Aerospace Technician
	35.2	35.2	35.0	34.6	34.3
	Fertilizer Producer	Waste Recycler	Aquaculture Farmer	Slaughter Processor	Chinese Cook
	34.3	34.2	35.0	34.2	33.6
Low Female Proportion	Railway Tech	Construction Technology	Carpenter	Electromechanical Repairer	Water Transport Operator
	17.6	17.1	16.6	16.5	16.1
	Steelworker	Police Officer	Road Builder	Agricultural and Forestry Machinery Operator	Decorator
	16.0	15.7	15.3	14.2	13.7
	Waterproofing Technology	Shipbuilder	Scaffolding Erector	Mason	Equipment Installer
	13.2	13.1	10.8	9.7	9.4
	Firefighter	Security Guard	Driller	Miner	Ship Commander/Pilot
	8.8	8.3	7.7	7.4	5.6

, the resulting 60 occupations were used as the core keywords for AI image generation and search engine image retrieval.

3.2. Data Collection and Image Acquisition

This study adopts the scraping audit method in the algorithm auditing approach, collecting image data from AI platforms and search engines. Algorithmic auditing resembles a natural experiment, serving as a specialized approach to uncover and diagnose biases, discrimination, and erroneous judgments within algorithmic systems [47]. Scholars classify it into types such as code audit, non-invasive user audit, scraping audit, sock puppet audit, and crowdsourced audit. Among them, scraping audit refers to the auditing conducted by researchers who do not interact with users and directly apply operations to the algorithm system to obtain output results. It has been widely applied in algorithm auditing research [2,48,49]. Given those multiple scholars, including Vlasceanu and Amodio [50], have confirmed that top-ranked images in search engines exhibit more pronounced biases, this study scraped the top 40 images from search engine results for 60 occupational keywords to capture evident algorithmic biases. Correspondingly, the AI platforms also generated 40 images using these same 60 occupational keywords. All the data were collected and completed by August 2025.

Given the fundamental differences between search engines and AI platforms, platform-specific input strategies were adopted to accommodate their distinct interaction logics. Search engines primarily retrieve and rank existing web content in response to concise keyword-based queries. Accordingly, relatively naturalistic search phrases were used to approximate ordinary occupational image searches. For example, using “translator” as the occupation, the search term was “solo work photos of translators.” In contrast, text-to-image AI models generate synthetic images conditioned on textual prompts. To improve the validity and consistency of gender identification across generated outputs, prompts included standardized instructions regarding image composition and subject positioning. Examples include: “Generate 4 distinct solo work photos of translators, all full-front views, showcasing professional work states in different settings.” These instructions were designed to reduce the occurrence of duplicated figures, non-frontal views, group portraits, or visually ambiguous subjects that could hinder reliable gender coding. Given that model outputs may be influenced by the gender pronouns and role semantics in prompts, Chinese syntax is better equipped than English linguistic structures to curb biased expressions in large language models to a certain extent [51,52]. Consequently, all prompts in this study were entered in Chinese to avoid assigning any identity or gender role to the model, thereby minimising the impact of syntactic and cultural differences on the model’s outputs as much as possible, and ensuring that any gender patterns stemmed solely from the model’s internalised schemas. Given the differences in prompt structure across various platforms, the aim is to accommodate the fundamentally distinct retrieval and generation mechanisms of search engines and generative AI, whilst maintaining sufficient consistency for the analysis of occupational gender representation. Using these two strategies, the top 40 images for each occupational keyword were retrieved from search engines, and 40 valid images were generated for each keyword via text-to-image models. After manually removing invalid samples, a total of 9600 raw image datasets were obtained from four platforms for analysis. Subsequently, the occupational gender representations generated by the retrieval-based system and the generation-based system were compared under conditions that met the platform query criteria.

3.3. Coding Rules and Analytical Methods

Based on feminist technology studies, this study constructs a cross-platform auditing framework to investigate how algorithmic systems differentially represent occupational gender characteristics. The framework integrates three core elements: a scraping audit method to identify bias in algorithmic systems; a tri-modal interpretive framework for visual text analysis; and a feedback loop model that elucidates the mechanisms through which bias is formed across technological, media, and sociocultural dimensions. This framework provides an integrated analytical approach for the systematic detection of generative and retrieval biases, thereby addressing research questions regarding the manifestations of bias, cross-platform differences, and their underlying mechanisms.

British scholar Gillian Rose proposed three interpretive approaches for visual materials: technological modality, social modality, and constitutive modality [53]. This study adopts this interpretive framework, with specific analytical steps outlined in

Table 2. Image Analysis Methods.

Modality	Content	Operation Method
Technological Modality	Mapping Technology	Acquire image data using AI platforms (Doubao, Jimeng) and search engines (Baidu, Sogou).
Social Modality	Economic, Political, and Social Relations in Images	Classify 60 occupations into 12 categories by social functions:	Education & Training: Preschool Teacher, Primary School Teacher, Special Education Teacher Healthcare: Nurse, Nursing Aide, Pharmacist, Traditional Chinese Medicine Physician Business & Finance: Accountant, Statistician, Securities Professional Scientific Research: Sociological Researcher, Agricultural Science Researcher Public Safety: Police Officer, Firefighter, Security Guard Manufacturing & Processing: Water Transport Equipment Operator, Ship Commander & Pilot …… (Note: Only partial categories are presented here due to space limitations)
Constitutive Modality	Semiotic Content in Images	Conduct manual coding and statistical analysis of the gender of figures in images.

. The technological modality refers to the distinct technologies used to acquire images: AI platforms and search engines. The social modality denotes the economic, political, social relations, and institutional practices reflected in images. The constitutive modality primarily encompasses the semiotic content of images, such as gender, age, demeanor and props of depicted subjects. To analyze the social modality, researchers further categorized 60 occupations into 12 types based on social function: education and training; healthcare; culture and media; business and finance; scientific research; public safety; social services; manufacturing and processing; agriculture, forestry, and fisheries; engineering and construction; transportation; and production operations. The study then examined patterns of gender representation bias in AI-generated and search engine images across these occupational categories. Within the compositional modality, the gender of individuals in images was specifically coded and statistically analyzed through manual coding. Given search engines’ inability to guarantee exclusively single-person photos in results and the current instability of AI image generation technology, occasional multi-person photos were encountered. Accordingly, the following coding rules were applied: if the image contained a single person, it was coded according to that individual’s gender. If multiple people of mixed genders were present, we adopted Kassarjian’s approach, focusing on the most prominent face in the multi-person image—that is, coding based on the gender of the central figure in the frame [54].

This study systematically examines gender representation biases across different professions on generative AI and search engine platforms. First, for the 40 images generated by each platform under each profession, we employ exact binomial tests to measure whether the observed number of female images significantly deviates from the actual female proportion in that profession [55]. To correct for multiple comparisons across four platforms, Bonferroni correction was applied [56], setting the statistical significance threshold at α = 0.05/240 ≈ 0.0002.

To assess overall gender bias at the platform level, we employed a one-sample proportion z-test, comparing observed proportions on each platform against weighted real-world proportions. Subsequently, a two-proportion z-test was used to compare the proportion of women depicted in images of a given occupation on one platform versus other platforms, testing for significant differences across platforms [57]. The study also incorporates analysis using the Quadratic Assignment Procedure (QAP) to examine the similarity in the structural distribution of female occupational images across the four platforms. QAP overcomes the independence assumption constraints of traditional statistical methods for network data through matrix transformation, enabling effective assessment of structural similarity between different matrices [58]. First, a 2-mode network linking occupations and data sources was constructed from raw data including the Sixth National Population Census statistics and data from Doubao, Jimeng, Baidu, and Sogou. This network was then transformed into a 1-mode network across platforms to reflect structural relationships in the distribution of female occupational image proportions. Subsequently, QAP correlation analysis was applied to the 1-mode network to identify structural homogeneity in occupational gender representation patterns across these four platforms.

4. Results

4.1. Overall Bias: Platform Representation Distortion Relative to Reality

In terms of overall gender distribution, significant differences exist among the four platforms, and both the trends in gender bias and the actual conditions exhibit systematic deviations (see

Table 3. One-Sample Z-Test Between Platform and Reality.

Platform	Total Female Count	Observed Proportion	Reality Proportion	Z Statistic	p-Value
Doubao	976	0.407	0.391	1.541	0.123
Jimeng	586	0.244	0.391	−16.769	0.000
Baidu	1039	0.433	0.391	4.123	0.000
Sogou	865	0.360	0.391	−3.143	0.002

). Z > 0 indicates that the overall female proportion observed in images on the platform is higher than the actual statistical data, and p < 0.05 signifies a significant deviation.

Figure 1 shows that among generative AI platforms, Jimeng exhibits a more pronounced imbalance in image gender distribution than Doubao. Specifically, “Jimeng > Doubao ≈ Real-world Situation.” Jimeng tends to generate images predominantly featuring males, while Doubao’s average overall gender ratio in images aligns more closely with real-world gender proportions. This may stem from general-purpose platforms receiving richer training data than specialized AI image generation platforms, consciously avoiding extreme biases during image generation. Consequently, Doubao’s generated images exhibit a gender distribution that more closely mirrors real-world proportions. This overall gender imbalance provides preliminary evidence of meaningful differences in the levels of generative bias between the two AI platforms.

Search engines exhibit representation patterns distinct from generative AI. Baidu and Sogou show similar trends in female representation within occupational images, both broadly aligning with statistical female distribution data. Among Baidu’s 2400 images, 1039 featured female subjects, representing 43.3%. Sogou featured 865 images with female subjects, accounting for 36%. The average female representation across 60 professions in reality was 39.1%. Baidu’s female proportion exceeded the real-world average, while Sogou’s fell below it. Therefore, the ranking of gender bias in search engines is “Sogou > Reality > Baidu.” Although Baidu and Sogou exhibit similar overall trends in gender bias, Sogou’s algorithm, due to its stronger inherent retrieval bias, reduces the visibility of women in professional roles.

4.2. Bias Pattern Analysis: Filtering Versus Generating Gendered Occupational Representations

Figure 2’s kernel density estimation curves reveal the overall distribution patterns of female proportions generated by each platform, compared against a baseline of real-world proportions. This enables the identification of concentrated trends and extreme values in bias, allowing for a more comprehensive assessment of bias patterns. Taking Jimeng as an example, its distribution exhibits pronounced dual extremes, with density peaks at both 0 and 1. This indicates an extreme tendency to categorize occupations absolutely as either “male-dominated” or “female-dominated.” In contrast, the female proportion in images generated by Doubao mostly clusters around the midpoint of 40–50%. Compared to AI, Baidu and Sogou exhibit more balanced distributions of female proportions across high, medium, and low ranges. Regarding the distribution density of low female proportions in occupational images, Sogou and Baidu share similar characteristics—both exhibit peaks. However, Sogou’s occupational images show higher density in medium female proportions, while Baidu’s images exhibit higher density in high female proportions.

Figure 3 provides concrete evidence of occupational image distribution patterns reflecting bias. AI platforms exhibit greater deviation from the baseline occupational gender proportions compared to search engines. In images generated by Jimeng, female-dominated occupations show systematically inflated levels of female representation, with eight occupations exceeding real-world proportions. Specifically, preschool teachers (93.3% female in reality), nurses (93.1%), childcare providers (89.2%), beauty/hairdressers (60.6%), and nursing aide (60.5%) are depicted as exclusively female, reaching 100% female representation in generated images. Notably, the discrepancy between Jimeng’s 100% female representation and the approximately 60% female proportion observed in real-world data for hairdressers and nursing assistants reaches nearly 40 percentage points. Conversely, traditionally male-dominated professions like engineers and programmers showed female proportions approaching zero suppressed by the algorithm to levels far below reality. This demonstrates that Jimeng exhibits dual extremes in occupational gender ratio data: extremely high female representation in professions with high female prevalence, and extremely low female representation in professions with medium-to-low female prevalence.

Doubao exhibits harmonizing characteristics. A comparative analysis of the female proportion data generated by Doubao across various occupations against real-world data reveals that in occupations with high, medium, and low female representation, Doubao’s female proportions are 54.9%, 43.3%, and 23.9%, respectively, while the actual figures are 68.2%, 36.6%, and 12.6%. While Doubao’s female representation in high-female-proportion occupations falls below reality, it exceeds actual figures in all other categories. The central peak in the kernel density map further confirms that Doubao creates an unrealistic “balance,” reflecting its effort to mitigate gender bias.

The distribution patterns of Baidu and Sogou align more closely with real-world data, both exhibiting multiple peaks that reflect the complex structure of occupations with high, medium, and low female representation in reality. For occupations with high female representation, Baidu and Sogou show average female proportions of 83.3% and 73.3%, respectively, while the real-world average is 68.2%. This indicates that in female-dominated occupations, search engines tend to amplify female representations and reinforce gender stereotypes. While Baidu and Sogou exhibit similar overall trends, notable differences emerge across specific occupations. In occupations with moderate female representation, both platforms display the most pronounced fluctuations. This suggests unstable search engine representations in professions with relatively balanced or neutral gender ratios. For instance, among funeral attendants, where women constitute 35.2% of the workforce, Baidu excessively emphasizes female representation, showing a female proportion as high as 92.5%. In contrast, for Chinese cooks, where women make up 33.6% of practitioners, search results are overwhelmingly male, with a female proportion of 0%. Such inconsistencies suggest that search engine representations may be driven more by algorithmic prioritization of specific keywords or visual cues than by faithful reflection of real-world data.

For occupations with lower female representation, Baidu and Sogou’s search result curves consistently fall below real-world levels. For instance, among waterproofing technicians, where women constitute 13.2% of the workforce, both platforms show zero female representation. In Sogou, female representation was also zero for carpenters (16.6% in reality), water transport operators (16.1%), steelmakers (16.0%), firefighters (8.8%), and drillers (7.7%), resulting exclusively male imagery. This indicates that search engines tend to further diminish female visibility when presenting traditionally male-dominated occupations, thereby reinforcing gender stereotypes.

The analysis reveals that AI models can creatively represent gender, with the general-purpose platform Doubao exhibiting pronounced “harmonization” biases and the specialized text-to-image platform Jimeng displaying extreme “polarization” biases. The recommendatory bias in search engines’ occupational gender representation manifests primarily as follows: in female-dominated fields, women are overrepresented, excessively reinforcing stereotypes about female occupations; conversely, in traditionally male-dominated fields, search engines fail to reflect women’s actual presence, showing a lower proportion of women than in reality, with multiple occupations even prioritizing male images. This “filtered reality” retrieval bias not only departs from actual occupational gender composition but may also affect perceptions of occupations through the overrepresentation of dominant genders and the underrepresentation of minority genders.

To examine this observation at the structural level, QAP correlation analysis was employed to assess the similarity of occupational gender structures across the four platforms (

Table 4. QAP Similarity Analysis of Occupational Gender Structures Across Platforms.

	Doubao	Jimeng	Baidu	Sogou	Reality
Doubao	1
Jimeng	0.484 ***	1
Baidu	0.549 ***	0.682 ***	1
Sogou	0.564 ***	0.684 ***	0.881 ***	1
Reality	0.618 ***	0.688 ***	0.863 ***	0.883 ***	1

Note: *** p < 0.001.

). The findings show that search engines exhibit a high correlation (r > 0.85) with real-world occupational gender structures, validating their “filtering” logic. In contrast, AI models show significantly lower structural correlation with reality (r ≈ 0.6), statistically demonstrating a tendency toward greater gender polarization in generated outputs. This indicates that search engines more closely align with the actual proportion of women across different occupations in reality than AI do in generating occupational female image distributions. Furthermore, the structural differences between the two search engines are relatively small (r = 0.881), while the structural differences between the two AI platforms are significantly larger (r = 0.484).

4.3. Differences in Algorithms and Visual Presentation Across Platforms

4.3.1. Differences in Algorithmic Bias Patterns: From Macro-Level Discrepancies to Micro-Level Variations

To assess differences in algorithmic bias patterns, the overall proportion of female practitioners across platforms was first compared at the macro level. As shown in

Table 5. Results of Platform-to-Platform Bivariate Proportion Z-Tests.

Platform 1	Platform 2	Observed Proportion 1	Observed Proportion 2	Z Stat	p Value
Doubao	Jimeng	0.407	0.244	12.015	0.000
Doubao	Baidu	0.407	0.433	−1.843	0.065
Doubao	Sogou	0.407	0.360	3.295	0.001
Jimeng	Baidu	0.244	0.433	−13.817	0.000
Jimeng	Sogou	0.244	0.360	−8.769	0.000
Baidu	Sogou	0.433	0.360	5.134	0.000

, significant differences in overall female representation were observed between most platforms, particularly with the AI platform Jimeng exhibiting extremely significant differences from all other platforms (p < 0.001), indicating divergent overall bias tendencies among different algorithms. To further examine the source of macro-level divergence, platform-to-platform two-proportion Z-tests were conducted for all 60 occupations (Figure 4).

As shown in Figure 4, statistical significance was evaluated by double proportion Z-tests with a significance threshold of p < 0.05. The position of each point reflects the relationship between the real-world female proportion for a given occupation (x-axis) and the female proportion generated by a specific platform (y-axis), while point color indicates whether that occupation differs significantly from at least one other platform. To minimize overlap caused by discrete proportional values, the study introduced minor positional fluctuations during visualization, which did not affect the fundamental statistical results. Red dots in Figure 4 indicate that gender representation for a given profession on a platform differs significantly from at least one other platform, identified in this study as a “point of contention.” White dots represent “points of consensus,” and the diagonal line indicates alignment with real-world statistics. This reveals, first, that contention is widespread. All platforms predominantly feature red points, indicating algorithms are far from unified in representing specific occupations’ gender compositions, instead generating substantial contradictory biases. Dispute patterns also vary. Jimeng’s extreme approach makes it a “dispute point” for the vast majority of occupations. Doubao’s “moderating” nature fosters its unique divergence pattern. In contrast, Sogou exhibits more “consensus points,” suggesting a filtering tendency that may approximate a neutral representation of online information. Disputed points often coincide with significant deviations from reality, indicating that algorithmic mechanisms may jointly contribute to distortions in gender representation.

4.3.2. Gender Disparities in Social Modalities

Sixty occupations were categorized based on social function using China’s demographic data, and differences in female proportions between real-world data and images on the four platforms were computed across 12 occupational categories. The core metric for calculating the discrepancy is “proportion of women in real data minus proportion of women in AI/ search engine images.” A positive difference indicates the platform generates fewer women than reality and more men than reality; a negative difference indicates the AI generates more women than reality and fewer men than reality. Next, the average difference across occupations within each social category was calculated to deeply interpret the bias characteristics of AI platforms and search engines in the social modality, as shown in Figure 5.

The data reveal contrasting bias patterns between Jimeng and Doubao across specific social modalities. Jimeng drastically underrepresented women in nearly all categories emphasizing elite authority or power-intensive professions, including Business & Finance, Scientific Research, Manufacturing & Processing, and Agriculture, Forestry & Fishery, while overrepresenting women in Education & Training. Doubao exhibited overall neutrality with relatively small deviations. Notably, both AI platforms underrepresented female practitioners in Business & Finance (Figure 6). In AI-generated images, males in these professions typically wear formal attire like suits and dress shirts. Their expressions convey confidence and composure, with poses focused on work. Even when faces are obscured, body shapes, clothing, and movements remain consistent. Some images also subtly imply hierarchical power dynamics through spatial positioning, suggesting the men occupy higher-status positions.

The filtering bias within search engines is complex. Baidu and Sogou exhibit similar bias patterns across most occupational categories. For instance, in traditionally male-dominated fields, including Manufacturing & Processing, Agriculture, Forestry & Fishery, Engineering & Construction, and Transportation, both platforms reinforce gender stereotypes by disproportionately featuring male imagery. However, biases diverge sharply in four categories: Business & Finance, Scientific Research, Public Safety, and Social Services, with Baidu tending to display female images and Sogou underrepresenting women. These divergences may stem from differences in recommendation algorithm design and training datasets, which in turn shape their distinct interpretations of occupational gender dynamics and resulting recommendations. Simultaneously, in fields like culture and media, both platforms exhibit relatively balanced representation, indicating that search engine biases are closely tied to societal perceptions of specific occupations.

4.4. Visual Comparative Analysis of Generative Bias Versus Retrieval Bias

4.4.1. Correlation of Bias Mechanisms

Although generative bias in AI image generation manifests differently from recommendation bias in search engines, they share certain similarities in the underlying mechanisms of occupational gender bias and its social impacts. At their core, both can be attributed to a mechanism of “algorithm black box—mass media—social cultural factors” (as shown in Figure 7). The cyclical feedback loop among audiences, media, and algorithms constitutes a complex adaptive system.

Inside the algorithm black box, both search engines and artificial intelligence models rely on real data sets for training, and thus inevitably absorb the long-standing gender biases and stereotypes in society [59]. Search engines dominate the information that the audience is exposed to through automated filtering, sorting, and organization mechanisms [60]. The generative bias of artificial intelligence models mainly stems from the imbalance of training data and the algorithm’s decoding learning of the data. Therefore, the gender bias narratives implicit in the mass media and other communication channels influence search engines and artificial intelligence models.

The training corpora of large models come from internet information resources [31], and the images generated by artificial intelligence are increasingly used for media dissemination and captured by search engines, further becoming training data for recommendation algorithms and generation models. Thus, the mass media not only constitute the data source of algorithm bias but also become an important amplifier for the dissemination and spread of bias [61].

The important causes of algorithm bias include the training process of input data, the process of platform algorithm development, and the actual application process of the algorithm. Therefore, it is necessary to deeply explore the internal operation mechanism of the “black box” and the platform development mechanism. However, the understanding of the technical system should not only remain at the technical level of operational logic, but also need to pay attention to its social and cultural impact at the practical application level. This circular amplification mechanism indicates that the visual presentation in the platform is not merely a technical output, but the digital reproduction of social cultural norms and gender order in the algorithm system. In the training and recommendation process of algorithms, they not only absorb the existing gender concepts in real society, but also further reinforce these stereotypes through continuous content generation and distribution. At the social and cultural level, after the audience is exposed to relevant media narratives, it may influence their behavior, attitude, and cognition; and the solidification of the public’s existing concepts will, in turn, shape the media production and algorithm training process, embedding genderized perspectives into the data.

Algorithms have now become a new power entity. At the platform ecosystem level, unlike the one-way dissemination of information by traditional media, search algorithms and generation algorithms mutually influence and interact with media and audiences, forming a data cycle structure that is mutually dependent. The discovery of the “systemic amplification” of occupational gender bias in this study further confirms that algorithm bias is not a single technical problem, but a result of the joint action of platforms, media, and social culture.

4.4.2. Visual Disparities in Bias Presentation

Gender bias in search engines and AI platforms exists not only quantitatively but also in presentation methods. As mentioned earlier, AI-generated images exhibit greater creativity and subjectivity in visual representation. AI models create entirely new images based on input instructions, where generative algorithms influence elements like character depictions and scene settings, often amplifying exaggerated occupational gender stereotypes. For instance, when generating images of “railway engineering technicians,” Doubao might depict male engineers as taller and more muscular, clad in professional workwear, wielding complex tools, and situated in more authentic industrial settings. Female engineers, conversely, may be portrayed as less prominent, wearing more ordinary attire, carrying fewer specialized tools, and positioned in less professionally defined environments.

Images generated by Doubao and Jimeng frequently share a common base face template, with generated characters exhibiting only subtle variations in expression, facial features, or hairstyle length. Currently, AI image generation is commonly used in artistic creation and design fields. The uniform presentation of faces may limit the diverse expression of professional imagery. Furthermore, unlike Jimeng, Baidu, and Sogou which predominantly feature Asian faces, Doubao not only emphasizes gender equality but also demonstrates a pronounced commitment to balanced diversity across nationalities and ethnicities. It avoids linking specific nationalities or ethnicities to particular professions, mirroring its gender equality approach. Together, these elements reflect Doubao’s balanced construction of professional image diversity across multiple dimensions.

5. Discussion

The analysis of the data in the previous text indicates that search engine algorithms exhibit retrieval bias, while AI models demonstrate generative bias. In the algorithm-dominated digital era, these biases represent the continuation of gender biases from the mass media age into the digital domain. Their core distinction lies in the power mechanisms shaping reality. Specifically, search engines use recommendation algorithms to “curate reality”: a curated presentation of actual conditions, thus amplifying gender stereotypes. AI models, however, exhibit greater generative capabilities. Through machine learning, they creatively shape the gender representations, demonstrating more complex and diverse tendencies. The connection between these biases lies in their joint formation of a symbiotic human–machine bias cycle. AI models and search engines learn from biased data and amplify these biases. AI-generated images also flow into search engines, influencing their biases. Audiences exposed to biased information further reinforce their own biases, creating a continuously worsening feedback loop of prejudice.

Feminist tech studies emphasize that gender and technology are co-produced, jointly shaping identities and social relations [62]. As digital cultural spaces, large language models and search engines bring together three key actors—platforms, audiences, and developers—who collectively reproduce gendered order under the impetus of algorithmic logic and market mechanisms. Seemingly neutral algorithmic technologies actually replicate imbalanced gender distributions by “labeling” gender differences and excluding diverse professional subjects. This further develops the “male gaze” in media communication into a “coded gaze” within algorithms [63].

The QAP analysis reveals that the consistency between occupational gender structures and real-world data across four platforms follows the order: Sogou > Baidu > Jimeng > Doubao. This indicates that there are significant differences in gender representation across different platforms. The search engine’s retrieval bias is mainly based on the existing information on the internet: data sources such as news, image libraries, and social media have long contained long-standing gender power relations, and the algorithm further reinforces gender bias through sorting and recommendation mechanisms [64]. However, the retrieval bias is not uniform internally. This study found that in professional fields such as business finance and scientific research, Baidu and Sogou exhibit completely opposite gender biases, which once again indicates that the “reality filtering” mechanism of search engines is influenced by the algorithm design and training data differences of each platform.

Compared to search engines, AI models exhibit more severe deviations from the reality. These deviations are not only continuations of socio-historical constructs but also outcomes of embedded values. Among them, the bias levels of the general-purpose AI platform Doubao and the specialized text-to-image platform Jimeng differ significantly, with distinct patterns of bias. Doubao’s deviations from reality stem from its internal algorithms’ pursuit of gender and racial equality. Jimeng’s biases, however, exhibit more extreme entrenchment in male-dominated professions, arising from its over-learning of mainstream data patterns, which exacerbates the visibility crisis for working women. These differences demonstrate that algorithms are not value-neutral technical tools; their design and application embed specific gender power relations and practical logic. The algorithmic value biases embedded within different platforms essentially represent the digital negotiation and reproduction of the social gender contract by various social actors. Doubao’s strategy can be viewed as an active attempt to reshape the gender contract through code intervention, while Jimeng’s logic subtly reinforces the existing traditional gender contract. Thus, the foundational impressions constructed over social history are amplified by material cultures like algorithmic technology, thereby reproducing social gender power. Despite operating within the same reality, the algorithmic mechanisms of generative AI and search engines differing in their foundational design yield distinct gendered professional images that reflect divergent underlying social relations, values, and objectives. However, due to the “active creation” nature of generative bias, its impact on social equity, particularly occupational gender equality, may prove more pronounced.

At the levels of dissemination and cognition, search engines and AI models jointly form a cyclic mechanism that connects “algorithmic black box—mass communication—social culture”, in which narratives about gender and social identity operate as a “metamessage” translated into computable data by machines devoid of embodied experience [65], which in turn shapes both recommendation and generative outputs. This data-driven perception model is prone to “machine hallucinations,” collectively constructing a “pseudo-environment” that shapes audience perceptions. At the cognitive effect level, according to “media priming theory,” when audiences encounter stereotypical content or concepts related to occupational gender, their judgments become susceptible to influence. Research on media portrayals of gender stereotypes also indicates that the more frequently audiences encounter such stereotypes, the more likely they are to develop entrenched beliefs about gender differences in social roles, status, and legitimacy, thereby reinforcing existing gender hierarchies in society [66]. Billions of people daily access information and make decisions through various interfaces [4]. Individuals often perceive the gender distribution and role positioning of different occupations based on information generated through channels like mass media, viewing this as a reflection of the real world and thereby forming fixed occupational gender perception patterns in their minds. The complexity and urgency of governing this phenomenon far exceed those of traditional media. Therefore, relying solely on algorithmic iterations or market competition to resolve these issues is insufficient. Proactive, diversified governance strategies must be implemented. Audiences also need to break free from “bias echo chambers” at the informational, cognitive, and behavioral levels to prevent societal perceptions from polarizing toward occupational biases, thereby upholding the baseline fairness of social norms [67].

This study compared the professional images generated by the platform with the statistical gender distribution of occupations, which helps to gain a deeper understanding of how gender representations are reproduced in the digital visual environment of China. The 60 occupations covered in this sample survey include 40 occupations with highly separated genders and 20 occupations with relatively balanced gender ratios. According to census data, the gender distribution structure of occupations with highly separated gender distributions shows stability over time, such as Preschool Teachers (female 93.3%), Nurses (female 93.1%), Miners (female 7.4%), and Ship Commanders (female 5.6%). This stable gender structure in occupations is influenced by the Chinese local policy systems, external gender role evaluations, and labor division [68,69,70], and is also closely related to gender role expectations in social culture [71].

For mid-level mixed-type occupations, there has been a certain degree of change in the gender distribution in recent years. The platform data observed in this study shows a gender imbalance that is often higher than the reflection level of historical benchmarks. For example, in the images generated by Jimeng, the proportion of female Funeral Attendants is 92.5%, while the sixth census data shows it to be only 35.2%. This indicates that the gender representation in the occupations on the platform cannot be fully explained by the real occupational structure, but is further influenced by the platform’s dissemination logic and algorithmic generation mechanism. Therefore, relying solely on technological iteration or market competition is not sufficient to solve the problem of algorithmic bias. It is necessary to combine multiple intervention paths such as platform governance, content review, and public media literacy education to avoid the continuous spread of occupational gender bias in the digital communication environment.

6. Conclusions

This study delves into the specific manifestations of generative bias in AI platforms (Doubao, Jimeng) and retrieval bias in search engines (Baidu, Sogou) within occupational gender contexts. These biases function not merely as discrete technological errors but as systemic outputs of the information ecosystem, reflecting, amplifying, and perpetuating gendered social structures through their embedded feedback mechanisms. Analyzing across three interpretive dimensions—technological, social, and constitutive—it compares the similarities and differences between generative and retrieval biases across the four platforms. Through the data analysis in the previous text, both AI models and search engines exhibit significant occupational gender bias. The retrieval bias of search engines (Baidu, Sogou) primarily manifests as a “selective amplification” of reality, reinforcing existing stereotypes through the “overrepresentation of dominant genders” and the “invisibility of marginalized genders.” In contrast, the generative bias of AI models reflects greater deviation from real-world statistical data exhibiting more complex and extreme forms of bias. The specialized platform Jimeng exhibits highly polarized gender distribution characteristics—specifically, amplifying stereotypical gender ratios by overrepresenting women in female-dominated occupations while underrepresenting them in male-dominated fields—while the general AI platform Doubao displays a “corrective” deviation that strays from reality in pursuit of gender balance. Ultimately, this study reveals the symbiotic and cyclical relationship between these two biases within the mechanism of “algorithm black box—mass media—social cultural factors.” Both types of algorithmic bias serve as the “digital mirror image” of sociocultural norms and as “technological tools” shaping societal cognition.

This study also has certain limitations. In terms of subject selection, it focused solely on four widely used AI platforms and search engines within China, without conducting in-depth research on other platforms. This may limit the generalizability of the findings. In addition, the structured design of the instructions may limit the naturalness of the model’s output, so the measurement results may not fully reflect the model’s actual performance and deviations under specific interaction conditions. Third, based on the Chinese context, we only classify images into two binary categories: male and female. We have not yet included non-binary genders or other gender identity groups. Future research could proceed in three directions: First, expand the scope of data collection to include more AI image generation platforms and search engines, encompassing both mainstream and niche platforms internationally and domestically. This broader sample analysis could distill more universal patterns of bias. Secondly, in regions that adopt a multi-gender classification system, it is advisable to go beyond the binary gender framework and further explore the characteristics of various gender identities presented in the algorithmic system. Third, future research can be based on the entire generation process of algorithmic bias, focusing on the pre-processing stages such as training data selection and algorithm model development for in-depth exploration. It aims to trace the root causes of algorithmic gender bias and further enrich and improve the related research system on gender bias in artificial intelligence algorithms.

The value of this research manifests in both theoretical and practical dimensions. Theoretically, it provides empirical analysis from a Chinese context for feminist tech studies, validating the “gendered” nature of algorithmic technologies through comparative analysis of search engines and AI platforms. Second, by comparing the operational mechanisms of retrieval bias and generation bias, this study proposes an algorithmic bias auditing framework at the methodological level, which is replicable and applicable across different platforms. This framework collects platform output data through scraping and auditing methods, and interprets the gender representation logic underlying the data using an interpretive approach based on the three modalities—technological, social, and constitutive modalities—and reveals the generation and evolution mechanism of bias in algorithmic technology, media, and social culture through a cyclic feedback mechanism model. Compared with previous case studies focusing on a single platform, this research constructs a transferable cross-platform research path, which is not only applicable to the analysis of occupational gender bias in the Chinese context but also provides methodological references for algorithmic bias research, platform auditing research, etc., in other countries, regions, and different types of platforms. Furthermore, by analyzing the specific differences in bias patterns across various AI platforms, it reveals the critical role of corporate values and algorithmic ethical design in shaping technological bias, shifting algorithmic bias research from macro-level sociocultural critique toward the meso-level of organizational design.

At the practical level, this research offers intervention insights for various stakeholders. Algorithmic gender bias is the result of dynamic interactions among the distribution of training data, algorithmic optimization logic, platform governance structures, and deeply entrenched gender norms. This emergent characteristic of information ecosystems highlights the limitations of addressing bias through isolated technical interventions alone, which are insufficient to break the cycle of systemic bias. Comprehensive governance strategies must be adopted that fully account for the feedback mechanisms among these interacting factors. For technology developers and platform operators, findings indicate that unsupervised learning alone inevitably replicates or even amplifies societal biases. Platforms must assume ethical responsibility for “algorithms for good,” implementing proactive interventions during data preprocessing, model training, and output stages. While Doubao’s “corrective” strategy is imperfect, it represents a positive direction for ethical exploration. Top-level design remains indispensable for steering algorithmic governance in the right direction. For policymakers and regulators, governing generative AI requires greater caution and foresight than search engines. Given the strong agency of algorithms in manifesting biases, mechanisms such as algorithmic ethics reviews and bias impact assessments should be considered to prevent erosion of societal consensus on gender equality. Furthermore, audiences must enhance digital literacy and critical thinking to guard against treating algorithm-generated images as objectively neutral knowledge. Vigilance is required against the “coded gaze” underlying such content, providing justification for integrating gender equality perspectives into science communication and technology education.