JCS 本刊论文 | 透视算法黑箱：数字平台的算法规制与信息推送异质性

Welcome to follow

The Journal of Chinese Sociology

2026年7月1日，The Journal of Chinese Sociology（《中国社会学学刊》）上线文章Peering into the algorithmic black box: algorithmic governance and information heterogeneity on digital platforms（《透视算法黑箱：数字平台的算法规制与信息推送异质性》）。

| 作者简介

刘河庆

华中科技大学社会学院教授

主要研究方向：计算社会学、人工智能与社会

梁玉成

中山大学社会学与人类学学院教授

主要研究方向：社会分层与不平等、计算社会学、中国社会转型

Abstract

Drawing on experimental and reverse-engineering approaches, this study situates algorithmic governance within the broader context of digital society and reflects on results of an experiment where multiple virtual accounts were constructed that engage in long-term, realistic interactions with digital platforms in order to penetrate the politicized “black box” of algorithms. It examines how algorithmic governance shapes the heterogeneity of information acquisition among users. Empirical findings reveal that in the digital era, algorithmic governance has become highly complex, sophisticated and concealed. At the information theme level, algorithms increase users’ exposure to diverse topical content. However, at the level of information semantics, algorithms reinforce the filter bubble effect, leading to the narrowing and enclosure of information flows. This study reveals how different individuals are situated by algorithms within relatively bounded positions in semantic vector spaces, and receive information confined to specific semantic dimensions.

Keywords

Algorithmic governance; Filter bubble; Thematic heterogeneity of information; Semantic heterogeneity of information; Digital society

Introduction

Algorithms are everywhere. In the digital age, as information proliferates exponentially, the space once available for individual operation, decision-making and choice has been increasingly replaced by algorithmic processes encoded in computer programs. Algorithms—together with the tools, services and platforms they enable—process massive and complex datasets that are beyond human capacity, and thus play an increasingly pivotal role across diverse domains. They have become essential to connecting, restructuring and mediating social relations within a digital society in which datafied systems and everyday social life are deeply entangled (Cheney-Lippold 2017). Common examples include the conveyance of personalized news and short-video feeds to social media users (Bail 2021) as well as the assignment and routing of food-delivery services through digital applications or apps (Chen 2020).

Given that algorithms have become integral to contemporary social life, there is an urgent need within the social sciences to examine them critically and empirically. It is important to understand how algorithms reorganize, mediate and mobilize preexisting social relations (Ruppert et al. 2013), and to clarify their potential societal implications. Take the widely used information-recommendation algorithms on digital platforms as an example; by automatically filtering and pushing information streams to users, such algorithms simultaneously influence and shape both individual patterns of information acquisition and the broader structure of public opinion and sentiment (Perra and Rocha 2019). Yet despite their central role in the digital age, there remains substantial scholarly disagreement over how algorithms concretely affect the diffusion and differentiation of information.

Several key questions emerge here: Do algorithms unlock access to high-quality, diverse content for users, or do they instead over-cater to individual preferences, continuously pushing homogeneous information that traps users within narrow filter bubbles? How does the dynamic interaction between large-scale, high-frequency algorithmic operations and individual behaviors position users within the information space? Do these processes further intensify informational isolation and differentiation among users? Answering these questions empirically is crucial – not only for clarifying how algorithms shape individuals’ access to information, knowledge production and ideological polarization, but also for understanding how algorithms, as a central technological artifact of the contemporary era, exercise power and shape social relations and realities in the digital age (Burrell and Fourcade 2021; Wang 2021).

Providing empirical answers, however, is far from straightforward. On one hand, the algorithms in question are typically highly complex, opaque and difficult to trace—operating within “black-box systems” (Mittelstadt et al. 2016). Platform companies rarely disclose detailed information about their algorithmic architectures or mechanisms, and direct access to proprietary data remains exceptionally challenging for researchers. On the other hand, even with partial access to algorithmic designs or operational details, it remains difficult to analyze their broader social implications. This difficulty stems from the fact that we now inhabit what scholars term “algorithmically infused societies” (Perra and Rocha 2019; Wagner et al. 2021), characterized by large-scale, high-frequency cycles of interaction among algorithms, training datasets, external constraints and vast populations of individual users. In Latourian terms, this constitutes a highly complex and dynamic network of heterogeneous actors (Latour 2005). Our object of inquiry, therefore, is not a simple or deterministic “algorithmic black box,” but a vast, networked system that continuously interacts with social reality (Seaver 2017). As a result, we cannot confine our discussion of algorithms to purely technical terms, nor can we accurately assess their social effects by analyzing only publicly disclosed principles or code fragments (Brown et al. 2021). These characteristics pose formidable challenges for empirical research in the social sciences.

To address these challenges, this study adopted an experimental and reverse-engineering approach. It employed virtual accounts as research instruments to simulate user attributes and observed their long-term interactions with algorithms and digital platforms. By doing so, this study seeks to enter what Amoore (2020) calls the “politicized space” of algorithms, thereby peering into the algorithmic black box and empirically examining how algorithmic governance affects the heterogeneity of users’ information acquisition. Specifically, using the T-platform—a highly algorithmically driven content-delivery platform—as a case, this study created 155 virtual user accounts with differentiated patterns of information-click behavior. Each account interacted continuously with the platform over a 25-day period according to its predefined preferences. This process generated a large dataset consisting of 233,973 news-feed pushes and over 2.94 million individual pieces of information. Based on these data, the study compares the structures and contents of information pushed to different virtual users along two analytical dimensions—thematic heterogeneity and semantic heterogeneity—in order to explore how algorithms influence individual information acquisition and inter-individual informational differentiation in the digital age, as well as the broader implications for governance.

Literature review

Algorithms as social power

In computer science, algorithms are typically understood as “control structures for performing assigned tasks” (Beer 2017). However, algorithms do not operate in a vacuum. Social science research has increasingly examined the social attributes and consequences of algorithmic systems (Qiu 2017), especially their role in constructing social order (Kitchin 2017). For example, credit-scoring algorithms classify individuals into risk categories that shape their access to loans, interest rates and financial opportunities (Kitchin 2017). Karen Young and Martin Lodge (2020) define algorithmic regulation as a decision-making system that governs behavior in a particular domain by learning from large-scale data to manage risks and modify behaviors toward predetermined goals. In other words, algorithmic regulation refers to the process by which algorithms categorize target populations based on massive data inputs (Amoore 2020) and then make automated decisions based on these classifications, thereby replacing traditional forms of human regulation. The social power of algorithms thus derives from this automated process of classification and decision-making (Burrell and Fourcade 2021; Thorson et al. 2021).

Unlike traditional human-based identification and classification, algorithmic recognition of “who we are” operates through multiple interpretive layers and involves thousands of possible classification targets (such as gender or hobbies). Any click or browsing behavior may become part of this classificatory decision process (Amoore 2020). In practice, clearly defined identities in the physical world are transformed into probabilistic, fragmented, and rapidly shifting identities within digital platforms. The main characteristic of algorithmic regulation, therefore, is its dynamic, modular form of control (Koopman 2019; Duan 2019). Within this process, how algorithms exercise power and by what standards they determine an individual’s likelihood of belonging to a particular category remain largely opaque—an inscrutable black box (Burrell and Fourcade 2021). Another manifestation of algorithms as social power lies in their reduction of inherently unquantifiable human differences into a single probabilistic output. Through probabilistic reasoning, algorithms may ignore, or forcibly standardize, subjective and contextual particularities, leading to classification errors and potential injustices in the treatment of individuals (Amoore 2020).

If the classification function of algorithms can be understood as the process of “knowing” human behavior, their decision-making function represents a large-scale process of “shaping” it. Existing studies of algorithmic decision-making and its consequences focus on three major aspects. First, the commercial logic of attention economies has led algorithm owners to employ addictive design strategies that constantly compete for users’ attention (Bakshy et al. 2015), resulting in the proliferation of fake news, clickbait and conspiracy theories (Bucher 2012; Sun and Liu 2018). Second, the precision but partiality of algorithmic regulation systems in large-scale information dissemination may reduce information diversity and the quality of public deliberation, thereby influencing knowledge production and the structure of public opinion (Chen 2022; Qiu 2022). Third, the black-box nature of algorithmic decision-making and its associated risks have left users with limited rights to explanation or recourse, while researchers find it equally difficult to evaluate the reasonableness and potential consequences of algorithmic standards (Perra and Rocha 2019).

In sum, the role of algorithms in automated classification and decision-making illustrates that algorithmic power in the digital age operates not through exploitation, but through exclusion (whether an individual is categorized) and invisibility (whether a certain type of information is pushed) (Lash 2009). However, existing research has largely described these phenomena externally, without deeply examining how algorithms specifically exercise exclusionary and invisible forms of power to shape social reality. In particular, within algorithmically infused societies, the long-term, dynamic interactions between algorithms and individuals and their deeper consequences, remain underexplored. Furthermore, much of the existing literature tends to conflate multiple types of algorithms within a single discussion. Empirical studies focused on specific algorithms are therefore essential for accurately understanding their differentiated effects across contexts and populations.

Algorithmic regulation

and the heterogeneity of

individual information acquisition

The most direct manifestation of algorithmic regulation lies in its dynamic control of information flows (Bail 2021). Through the targeted delivery of content, recommendation algorithms dominate the diffusion and circulation of information in the digital era (Bucher 2012; Bakshy et al. 2015; Bail 2021). Yet significant debate persists regarding how algorithms exercise power through exclusion and invisibility, and how these mechanisms affect individual information acquisition and inter-individual differentiation. The literature can be broadly divided into two lines of inquiry: “over-individualization” and “classificatory power.”

The over-individualization perspective argues that algorithms create micro-information environments for each user based on browsing histories and other personal characteristics. As algorithms continuously interact with individuals, these environments become increasingly homogeneous, reducing opportunities for users to encounter diverse information (Bakshy et al. 2015). This phenomenon is often described as the “information cocoon,” “echo chamber” (Sunstein 2008), or “filter bubble” (Pariser 2011). Pariser (2011) contends that platforms such as Facebook and Google use algorithms to infer and deliver content that aligns with users’ presumed interests, thereby constructing filter bubbles that cater to and amplify personal preferences. Users may ultimately be exposed only to biased and fragmented depictions of the world, intensifying informational isolation and ideological polarization.

The classificatory power perspective, by contrast, emphasizes the loss of individual agency under algorithmic regulation. Algorithms push content based on predicted categories, rendering individual preferences and intentions largely irrelevant. In digital societies, multidimensional categorization and indexicalization have profound implications for information access (Cheney-Lippold 2017; Amoore 2020). Methodological individualism has been replaced by indexicalization: individuals become mere data points within categorical codes, and personal uniqueness or subjectivity loses significance (Cheney-Lippold 2017). The essence of recommendation algorithms lies in transforming user data into multidimensional datasets through labeling and profiling, which enable targeted content delivery, population management, and behavioral guidance. The information an individual receives is thus based not on personal intention, but on what algorithms infer that “people like you” would find interesting (Duan 2019; Han 2019).

These two perspectives highlight possible effects of algorithmic regulation on individual information access and inter-individual differentiation. Yet, in actual algorithmic practice, it remains unclear which mechanism—over-individualization or classificatory power—predominates, whether algorithms indeed cause and reinforce filter bubbles, and what types of informational disparities emerge among individuals (Thorson et al. 2021; Bail 2021; Ge et al. 2020; Chen and Wang 2019). Existing empirical studies have yielded conflicting findings and remain fragmented (Shi et al. 2022). Some find that algorithms amplify individual preferences, narrowing informational exposure and increasing divergence between users (Bucher 2012); while others argue that the threat of filter bubbles has been overstated and that personalization effects are smaller than often assumed (Nechushtai and Lewis 2019). Kjerstin Thorson and colleagues (2021) discovered that users categorized by algorithms as politically interested were more likely to receive political content than users who self-reported such interests. Similarly, Jinghong Nie and Jiazi Song (2020) found in their study of health-information acquisition that users must actively exercise agency to access content of interest, contradicting the algorithmic promise of automated precision.

This paper argues that the theoretical and empirical contradictions above stem largely from conceptual compression and a lack of dimensional differentiation in prior studies. Much of the existing literature either discusses information in overly general terms or focuses on single dimensions (Lash 2009; Cheney-Lippold 2017; Amoore 2020). Information itself is multifaceted, encompassing dimensions such as source, theme and semantics. Moreover, digital platforms operate under the combined influences of commercial incentives, state regulation and public attention (Lv et al. 2022; Zhao 2022), constantly balancing between accuracy and diversity in information delivery (Helberger et al. 2018). The rise of deep learning has enabled platforms to capture increasingly subtle semantic features beyond basic thematic or source-level distinctions (Liu 2019), allowing for more fine-grained control over information dissemination to meet the varying needs of platforms, users and regulators. In this context, distinguishing between different dimensions of information—particularly by comparing how users differ in their exposure to coarse-grained thematic information versus fine-grained semantic information—can help clarify how algorithms exercise social power through the precise control of information flows and what social consequences such control entails.

In summary, existing research remains limited in three key respects. First, most studies analyze algorithmic effects on information heterogeneity using only a single dimension of information, typically thematic (Thorson et al. 2021), with few integrating multiple dimensions. Next, the real-time and dynamic nature of algorithms must be recognized; long-term continuous observation, rather than one-time snapshots, is necessary to capture their true effects. Finally, existing studies often rely on self-reported data or browsing histories, which confound algorithmic behavior with user preference. To overcome the endogeneity of user clicking behavior, an ideal research design should fix user actions and collect the complete set of information recommendations that algorithms deliver to users.

Research design

Data collection and processing

T Platform, one of the largest information-distribution platforms in China, relies entirely on recommendation algorithms to deliver content automatically across more than one hundred vertical domains, including technology, sports and current affairs. According to the platform’s publicly available descriptions of its recommendation principles, its algorithmic system primarily operates on three dimensions: user behavior features, content features and contextual features. It combines collaborative filtering with deep neural networks and other methods for information recommendation. In practice, the algorithm places particular emphasis on factors such as relevance, context, popularity and collaborative similarity (Liu 2019). Beyond the complexity of its input parameters and the diversity of features considered, another defining characteristic of T Platform’s recommendation algorithm is its real-time adaptability. The algorithm continuously retrains and updates its parameters online in response to changes in user behavior, content attributes, and environmental conditions, thereby enabling real-time, dynamic information delivery.

Building on these operational principles, this study sought to account for user, content and environmental characteristics as well as the algorithm’s real-time variability. To that end, we created twelve user groups (a total of 155 virtual accounts), each designed with distinct thematic clicking preferences, to interact continuously with the platform over time. This setup enabled the collection of large-scale data on algorithmically recommended information flows across different simulated users.

The research design followed four main steps:

Volunteers were first recruited to register accounts on the platform. After registration, the researchers configured each virtual account’s information preferences and clicking behaviors. The use of virtual accounts allowed for precise control of parameters according to the research design.

Then a preliminary data crawl was conducted using an information-feature collector to obtain the platform’s labeling and classification schema. Drawing on the platform’s annual user-behavior reports—which classify users by gender, age and city tier and analyze corresponding information-click preferences—we divided the 155 virtual accounts into twelve groups. Each group was assigned a distinct pattern of clicking preferences to simulate differentiated user profiles when engaging with real-time information feeds.

Table 1 reports the preferences of each group. Group 1 served as the random-control group, consisting of twenty accounts. When presented with a feed of approximately 14–15 information items, each account randomly clicked on 30 percent of items regardless of category, then refreshed to receive the next batch of information (see Fig. 1 for the process). Groups 2 through 9 simulated idealized user types based on platform reports. For instance, Group 2 represented male users from first- and second-tier cities in older age brackets, who clicked with 90 percent probability on items labeled as politics, finance, technology, nature or automobiles. Group 3 represented male users from third-tier and lower cities in older age brackets, who clicked with 90 percent probability on items labeled as society, military, history, law, health or world news. Groups 10 through 12 were extreme test groups, each clicking exclusively on a single type of labeled content.

It is important to note that this experimental design did not aim to fully replicate real-world user–platform interactions. Such replication would be unrealistic given the immense complexity of both user and platform parameters. Rather, the purpose was to examine how algorithms affect the heterogeneity of user information acquisition. By fixing users’ information preferences and clicking behaviors and then continuously observing changes in the content recommended to them, this study minimizes the endogeneity inherent in user click behavior, allowing for a more accurate assessment of algorithmic effects.

Figure 1 illustrates the overall data-collection process. Each virtual account first received a news feed (with a unique feed ID) containing multiple items, each labeled with an ID, title, summary, date and tag information. The account then selected whether to click on each item based on its pre-set preferences, generating a record of click behavior, and subsequently refreshed the screen to obtain a new feed—repeating this process continuously. In total, 155 virtual accounts interacted with the platform for 25 consecutive days at fixed time intervals according to their designated clicking patterns, and all information-flow data were saved.

Ultimately, the study produced a multi-level dataset containing 233,973 news-feed instances and their basic attributes (including feed IDs) as well as over 2.94 million individual items of information with associated features such as titles, summaries and tags.

Data analysis strategy and methods

Drawing on prior studies and the operational logic of the platform’s algorithms, this study analyzes the heterogeneity of algorithmic information delivery across two dimensions: information theme and information semantics.

Information-theme dimension

We first measure thematic heterogeneity by comparing the distribution and diversity of information themes pushed to different user groups and by computing the entropy index of information flows (Zhang et al. 2017). (1) In terms of group differences and longitudinal variation, if the algorithm indeed reinforces the filter-bubble effect, we would expect that, within the overall recommendation outcomes, each group would increasingly receive a higher proportion of information in the categories it initially preferred, and that the distributions of thematic content across groups would diverge as the experiment progresses. (2) The entropy index is used to assess the structural diversity of information categories within the feeds received by each group. If the algorithm amplifies the filter-bubble effect, we would expect groups with a broader range of preferred themes (such as Group 2) to exhibit higher entropy values (i.e. greater thematic diversity) than groups with narrowly defined preferences (such as Group 10).

Information-semantic dimension

Compared to thematic analysis, measuring heterogeneity at the level of information semantics is more complex. According to the platform’s published algorithmic principles, deep learning models convert high-dimensional user and content features into low-dimensional real-valued vectors, recommending information by comparing distances between user and content vectors. To analyze heterogeneity in this semantic space, the distance and dynamic variation between user and content vectors must be measured. Accordingly, this study employs a document-vector model (Doc2Vec) to model over 2.94 million pieces of information collected from the platform. Doc2Vec extends the word-vector model (Word2Vec) to sentences, paragraphs, documents or categories (Le and Mikolov 2014). By introducing document-level variables into lexical contexts (Rheault and Cochrane 2018), the model predicts words within documents and represents each document or category as a dense vector. Using the combined titles and summaries of each pushed information item as the corpus, a Doc2Vec model was trained with a window size of 5, a minimum word frequency of 10, and 20 training iterations, resulting in a 200-dimensional real-valued vector representation for each information item. After concatenating these 200-dimensional vectors with other item-level features, we applied principal component analysis (PCA) and multilevel fixed-effects models to examine heterogeneity in the semantic dimension of information acquisition across users. If the algorithm reinforces the filter-bubble effect at the semantic level, we would expect:

(1) Within-group distributions of user accounts in the semantic vector space to be relatively clustered rather than random;

(2) Significant differentiation between groups in the positions of their recommended information within the semantic space; and

(3) An expanding divergence among groups over time as their interactions with the platform deepen.

Empirical analysis

based on thematic heterogeneity

of information

Distribution of

pushed information themes

across user groups

This section first analyzes the overall distribution of information themes pushed to the twelve user groups after extended interaction with the platform. Table 2 reports the six most frequently pushed themes for each group and their respective proportions.

As designed, Group 1 serves as the random-control group. As shown in Table 2, political news accounts for the largest share of Group 1’s pushed content, at 27.98 percent. The next five most common themes are society, entertainment, world, history and finance, accounting for 7.50 percent, 7.31 percent, 5.16 percent, 5.15 percent and 4.62 percent, respectively. The proportion of political news received by Group 1 is notably higher than that of other groups (except Group 2). Although Group 1 had no specific thematic clicking preferences, political news constituted a relatively high proportion of the initial information feeds during the experiment (see Table 1). Given that this group randomly clicked on 30 percent of items per feed, the probability of clicking on political news was therefore correspondingly higher.

Table 2 also presents the thematic distribution of pushed content for Groups 2 through 12. Each of these groups was assigned fixed thematic preferences. In this table, preferred themes appearing among the top six categories are indicated in bold. The results show that when accounts are set to prefer particular topics, the platform indeed increases the volume of information in those categories over time. For example, in Group 3, the shares of its preferred themes—society (11.51 percent), world (6.98 percent), and history (6.21 percent)—are all higher than the overall group averages. However, the platform’s recommendation algorithm does not appear to overemphasize user preferences to the point of excessive reinforcement. Take Group 10 for example; as an extreme test group that clicked exclusively on “history,” the final proportion of historical news in its pushed information ranked fourth (7.55 percent), only slightly above the cross-group average (6.16 percent), and far from being dominant. Moreover, except for Groups 2 and 9, the theme with the highest proportion of pushed information in each group was not the group’s initially preferred category. This suggests that the algorithm does not fully determine information delivery based solely on user clicking behavior. Finally, despite differing preferences, each group’s pushed information themes show a high degree of overlap. Among the twelve groups, eight received “politics” as their top category and four received “entertainment.” In other words, differences in clicking preferences did not produce substantial divergence or differentiation in the thematic structure of recommended information.

Although these results do not directly negate the possibility of filter-bubble effects, they indicate that the platform’s recommendation algorithm does not exhibit the kind of excessive preference amplification emphasized in some prior studies. The threat of filter bubbles appears overstated (Nechushtai and Lewis 2019), and the effects of personalization are smaller than commonly assumed.

Longitudinal changes

in thematic information delivery

This subsection further examines how the distribution of thematic information evolves over time across groups, thereby revealing the dynamic interactions between the platform’s recommendation algorithm and user information acquisition. Figure 2 uses “politics” as an example, showing the temporal variation in the proportion of political news pushed to each group. The upper and lower panels of Fig. 2 respectively present the results for Groups 1–6 and Groups 7–12. The horizontal axis indicates the number of days since the start of the experiment, and the vertical axis represents the share of political news among all items pushed to each group.

As shown in Fig. 2, on the first day of the experiment, the proportion of political news pushed to each group was roughly similar—around 10 percent. Differences emerged thereafter. First, for groups with high political-click volumes (Groups 1 and 2), the algorithm rapidly increased the proportion of political content within a short time. However, as time progressed, the proportion of political news in these groups fluctuated and eventually declined, rather than remaining high or continuing to rise. Second, for Groups 3 through 12—many of which showed no political preference—political content consistently constituted a certain proportion of their feeds throughout the study. The variability across groups was greater in the early stages and smaller later on.

External content trends also influenced users’ exposure to political information. Around Day 18 of the experiment, the proportion of political news declined sharply across all groups, reaching a low point by Day 20. According to third-party popularity indices, a major celebrity scandal began trending on Day 18 and peaked on Day 20. This suggests that content “popularity” or “heat,” as incorporated into the algorithmic design, directly affects information distribution. At specific times, the algorithm prioritizes trending or popular content over predicted user preferences or categorical relevance.

Analysis of information diversity

based on the entropy index

To further evaluate the diversity of pushed information categories, we computed the entropy index for each information feed on a daily basis. Higher entropy values indicate greater thematic diversity within a given feed. We then used a two-level fixed-effects model to examine how the diversity of pushed information varied across groups and over time.

Table 3, Model 1, reports the effects of control variables, group variables, and date variables on the entropy index. Results show that Groups 1, 2 and 3 received less diverse information than other groups, while Groups 7, 8, 11 and 12 received more diverse content. Comparing these results with the groups’ clicking preferences reveals that having multiple preferred tags (e.g., Group 2) does not necessarily lead to greater diversity. In contrast, Groups 11 and 12—each with narrow preferences for rare categories—showed the highest entropy. Because their preferred tags represented low-probability categories (0.8 percent and 0.6 percent of the baseline distribution, respectively), their clicking patterns increased the probability of exposure to these categories, thus raising the entropy index.

Empirical analysis

based on semantic heterogeneity

of information

Differences among

individual accounts and

groups in the semantic vector space

This section examines information heterogeneity at a deeper semantic level. Using the Doc2Vec model, vector representations of the information (titles and summaries) pushed by the platform were trained. For each individual account and date, a 200-dimensional real-valued vector was generated. In Fig. 3, the average vector for each account was calculated followed by an application of principal component analysis (PCA) to project these account-level vectors onto a two-dimensional semantic space. Each data point in Fig. 3 represents an account, color-coded by its assigned group. In simple terms, the first principal component captures the largest variance among accounts’ semantic representations – that is, the major differences in the semantics of information received by different accounts (Rheault and Cochrane 2018). This subsection focuses primarily on variations both within and between groups along the first principal component.

Within-group distributions along the X-axis reveal that account placements are not random. Accounts within the same group tend to cluster closely together, indicating that the information pushed to members of the same group is semantically similar. In contrast, there are clear semantic distinctions between groups: Groups 1 and 2 are positioned toward the right end of the X-axis, whereas Groups 7 and 9 are located toward the left.

To visualize these differences more intuitively, we averaged the account-level scores within each group, as shown in Fig. 4. The results reveal that, along the first principal component representing semantic variation, Group 2 lies at the far right of the axis while Group 11 lies at the far left; these two groups exhibit the largest semantic distance. Furthermore, apart from the random-control group (Group 1) and the three extreme test groups (Groups 10, 11 and 12), all remaining groups form two distinct clusters along the X-axis. Groups 2, 4, 6 and 8 fall on the positive side, whereas Groups 3, 5, 7 and 9 fall on the negative side. Together, Figs. 3 and 4 indicate that the semantic distribution of pushed information across accounts is structured rather than chaotic. Within-group information content is relatively homogeneous, whereas between-group differences are pronounced.

Differences across groups

within sub-semantic spaces

The semantic differences observed above could potentially be driven by variations in information themes. To account for this, a more granular analysis within specific thematic subspaces is implemented. We selected pushed information belonging to six major themes—politics, society, entertainment, world, history and finance—and trained separate Doc2Vec models for each subset. For every account and date, we obtained a 200-dimensional vector within each sub-dataset. Following the same steps as in the previous subsection, we applied PCA and averaged the results by group. Figure 5 presents the relative positions of groups across these six sub-semantic spaces.

The results in Fig. 5 show that semantic differentiation persists even when focusing on specific thematic subspaces. Interestingly, the patterns of differentiation are remarkably consistent across all subspaces. With the exception of the random-control and extreme test groups, the positions of the other groups on the X-axis mirror those observed in the overall semantic space. In every subspace, Group 2 consistently appears at the far right, while Groups 7 and 9 consistently appear at the far left, forming a stable and regular pattern of group differentiation. Taking the “society” theme as an example, both Group 3 and Group 5 have explicit preferences for social news, and their positions in the social subspace are indeed close to each other. Other groups, meanwhile, are neither tightly clustered nor far removed from these two groups, but rather display a distribution pattern similar to that seen in the overall semantic space and in other thematic subspaces. Even groups that share a thematic preference for social news (Groups 1, 2, 7 and 9) receive content that is semantically distinct from one another.

The consistent distribution across sub-semantic spaces suggests that, at the level of information semantics, the algorithm reinforces the filter-bubble effect. Different groups receive information that is semantically differentiated and stratified. Although, at the thematic level, the platform continues to push some information outside users’ preferred categories, thereby increasing exposure to diverse topics and reducing thematic-level heterogeneity. At the semantic level the algorithm learns and estimates deep semantic preferences from user clicking behavior and delivers content accordingly. As a result, semantic differentiation across user groups becomes pronounced and consistent across subspaces. In other words, at this deeper semantic level, information delivery becomes narrowed and enclosed. Different user groups appear to be positioned by the algorithm within relatively fixed locations along the “semantic spectrum,” each confined to receiving particular kinds of political, social, entertainment, world, historical or financial content, respectively.

Further analysis using

a two-level fixed-effects model

Building on the account-by-date vector representations used above, this subsection focuses on trained document vectors at the level of individual pushed items to test the robustness of the prior findings and to analyze the dynamic evolution of semantic differentiation across groups. Using the Doc2Vec model, we trained a 200-dimensional real-valued vector for each individual piece of pushed information, then extracted the first principal component from the PCA results. A two-level fixed-effects model was then applied to examine the effects of group and date on the semantic dimension of pushed content.

Table 4, Model 1, reports the effects of control, group and time variables on the semantic component. Using Group 8 as the reference category, the results show that Groups 2 and 4 have significantly higher scores on the first principal component than Group 8, while Group 6 scores slightly lower. Conversely, Groups 3, 5, 7 and 9—all located on the left side of the X-axis in previous analyses—have significantly lower scores than Group 8. These results reaffirm the earlier findings: semantic differentiation among groups is clear and follows a stable pattern.

Model 2 introduces interaction terms between group and date. The main effects suggest that, in the early stages of the study, differences among groups were present but less stable than those described above. The interaction effects reveal that as user–algorithm interactions deepened, the groups gradually developed the clear, stable differentiation patterns identified earlier (including those in Model 1). Specifically, compared with Group 8, the scores of Groups 2, 4 and 6 on the first principal component increased over time, while those of Groups 5, 7 and 9 decreased (Group 3 was already lower at the outset). Over time, the groups converged into two distinct clusters within the semantic space: Groups 2, 4, 6 and 8 on the higher side of the component and Groups 3, 5, 7 and 9 on the lower side. This finding is particularly significant because it indicates that algorithms do not initially assign users to fixed semantic positions. Rather, through continuous interaction with user clicking behavior, algorithms dynamically reposition users within the semantic spectrum. Over time, this process produces increasingly stable and pronounced patterns of semantic differentiation.

Robustness check

To test the robustness of the findings regarding semantic heterogeneity, we extracted the top 50 principal dimensions from the high-dimensional document vectors trained in the previous Doc2Vec models. Using cosine similarity, we measured both inter-group and intra-group semantic similarity. Figure 6 shows that clear patterns of semantic differentiation across groups persist, consistent with earlier results. We also computed the average within-group semantic similarity for each group and assigned these values as node weights. As shown in Fig. 6, the average within-group semantic similarity is consistently high, meaning that semantic differences within groups are significantly smaller than those between groups.

Figures 3 through 5 in this study were based on document vectors trained at the account-by-date level. To verify the robustness of these results, we also trained document vectors at the level of individual pushed items. The results were highly consistent with earlier analyses, showing no significant deviations.

Conclusion and discussion

We now live in a society deeply permeated by algorithms—one in which algorithmic intervention poses profound challenges to social science research. Confronting the difficulties of algorithmic opacity, complexity and traceability, this study sought methodological innovation in both research design and data generation (Wagner et al. 2021). By creating multiple virtual accounts with distinct clicking preferences and enabling them to interact continuously with a major platform, we collected large-scale data on algorithmic recommendations and conducted an empirical analysis of heterogeneity across two dimensions: thematic and semantic.

The observed differentiation across groups in both dimensions demonstrates that algorithmic power in the digital age has become increasingly subtle (Lash 2009). Algorithmic regulation now exhibits traits of high complexity, refinement and concealment. At the societal level, competing demands—from platform profitability, external regulation and public opinion—intersect with advances in algorithmic capacity to capture deep semantics. This results in a pattern whereby the information pushed to users appears thematically diverse but semantically narrow. On one hand, the T platform’s use of popularity-based recommendation logic increases users’ opportunities to encounter diverse thematic information. This reduces concerns that algorithms necessarily intensify filter bubbles or restrict exposure to diverse topics, aligning instead with regulatory and public expectations. On the other hand, the empirical results reveal that, at a deeper and more concealed semantic level, the T platform engages in a form of fine-grained filtering of high-popularity information. By capturing and estimating users’ semantic preferences—however predicted rather than explicitly expressed—the algorithm progressively fixes individuals at specific positions along the semantic spectrum. Over time, each user is pushed information corresponding to a particular semantic dimension, leading to more subtle but real narrowing of information access and growing informational segregation between users.

These findings have several implications. First, research on algorithms must take into account multiple dimensions of information. Apparent contradictions in prior empirical studies can often be reconciled by disaggregating and aligning the analytical dimensions of inquiry. Second, as platforms increasingly integrate deep learning and related technologies, their operational logic is shifting from a traditional “classification logic”—based on coarse, static categories—to a “distance logic,” which governs information through fine-grained, real-time semantic proximity. This transformation makes algorithmic control over information flows more complex, refined and invisible. Comparing users’ access to information across both thematic and semantic dimensions provides insight into how algorithmic regulation exercises social power through the subtle management of information streams, thereby illuminating the potential social consequences of this mode of control.

The characteristics of algorithmic power identified here also imply that neither publicly disclosed algorithmic principles nor isolated user interviews are sufficient for analyzing how algorithmic power operates. To address this methodological challenge, this study draws on the approaches of algorithmic auditing in social science (Brown et al. 2021) and computational experimentation in computer science (Wang 2007). By using virtual accounts as research instruments to engage in long-term, automated interaction with real-world data environments, this study explores the extended temporal dynamics of algorithm–society interactions and their consequences. Such an approach offers social scientists a new methodological interface for examining the black-boxed technological world, enabling a closer entry into what Amoore (2020) calls the “politicized space” of algorithms. It thus provides both a novel data-collection strategy and a methodological pathway for empirically evaluating the role and social impact of algorithms in decision-making processes.

As an exploratory empirical study of algorithms in the social sciences, this research also has several limitations that invite further investigation. Future studies should incorporate more parameters, finer-grained thematic categories, and a wider range of digital platforms to deepen our understanding of algorithm–user interaction. Because data availability constrained this study, we defined user clicking preferences only at a coarse thematic level. Incorporating more granular behavioral parameters would enhance the precision of future analyses. Also, this design did not include behavioral indicators such as “likes” or “dislikes,” which influence the platform’s content delivery. Future research could incorporate these factors to observe corresponding changes in recommendation outcomes. Third, while the document-vector models used here effectively map group-level positions within overall and thematic semantic spaces, they do not yet clarify the precise nature of the semantic differences captured by the first principal component. Further research on the interpretability of such complex vector models is therefore warranted.

免费阅读并下载全文

引用本文

Liu, H., Liang, Y. Peering into the algorithmic black box: algorithmic governance and information heterogeneity on digital platforms. J. Chin. Sociol. 13, 13 (2026). https://doi.org/10.1186/s40711-026-00264-4

https://link.springer.com/article/10.1186/s40711-26-00264-4

文章仅为作者观点，不代表本刊立场

更多相关文章

以上就是本期JCS推文的内容啦！

定期查收讲座、征文信息/趣文推荐/热点追踪/主题漫谈

学术路上

JCS陪你一起成长！

关于JCS

《中国社会学学刊》（The Journal of Chinese Sociology）于2014年10月由中国社会科学院社会学研究所创办。作为中国大陆第一本英文社会学学术期刊，JCS致力于为中国社会学者与国外同行的学术交流和合作打造国际一流的学术平台。JCS由全球最大科技期刊出版集团施普林格·自然(Springer Nature)出版发行，由国内外顶尖社会学家组成强大编委会队伍，采用双向匿名评审方式和“开放获取”(open access)出版模式。JCS已于2021年5月被ESCI收录。2023年，JCS在科睿唯安发布的2023年度《期刊引证报告》（JCR）中首次获得影响因子并达到1.5（Q3）。2025年JCS最新影响因子1.3，位列社会学领域期刊全球前53%（Q3）。2025年JCS的CiteScore分值为3.4，在社科类别的286本期刊中排名第75位，位列同类期刊前27%（Q2）。

欢迎向《中国社会学学刊》投稿！

Please consider submitting to The Journal of Chinese Sociology!

官方网站：

https://journalofchinesesociology.springeropen.com