Social Science Computer Review

With the growing prevalence and accessibility of AI companions, contemporary women are forming relationships with virtual partners. It is important to examine the relational, social, and gender-related implications of this phenomenon. Our research sheds light on the complex power dynamics in young urban Chinese women’s engagements with the AI partner Replika. By analyzing 342 relevant posts from an online chatbot community, guided by Foucault’s concept of power, we uncover inherent and actively exercised power dynamics in three typical user-bot relational pairings: customer-product, human-machine, and woman-man/woman-woman. Following the Foucauldian theories, we analyzed the interplay of truth, desire, knowledge, and power. We discovered three specific neoliberal subjectivities that both free and constrain female users in their engagement with Replika. While these female users challenge traditional gender norms through erotic roleplays with Replika, Generative AI introduces potential risks of sexual harassment and gender bias channeled from the extensive online world through AI partners to individual users.

This paper provides a practical example and guide on how to augment or replace human coders with Large Language Models (LLMs) during content analysis. We demonstrate this by replicating and extending an influential study on environmental communication. Our setup, running locally on consumer-grade hardware, makes it feasible for university researchers operating within typical computational and legal constraints. We validate the LLM’s performance by replicating the original study’s codings, scaling the analysis to cover a tenfold increase in articles, and extending the LLM’s application to a comparable German-language corpus, comparing these results to human expert coders. We offer guidelines for instructing LLMs, validating output, and handling multilingual coding, presenting a replicable framework for future research. This paper is intended to systematically guide other researchers when integrating LLMs into their workflows, ensuring reliable and scalable coding practices. We demonstrate several advantages of LLMs as coders, including cost-effective multilingual coding, overcoming the limitations of small-sample content analysis, and improving both the replicability and transparency of the coding process.

Civil unrest, encompassing protests and riots, is an increasing global concern, with incidents rising at an alarming rate, a trend that has been observed in South Africa over the years. This issue is particularly pronounced in today’s social media era, where platforms like ‘X’ (formerly Twitter) serve as powerful tools for mobilization. This raises the question: What factors drive civil unrest, and how can machine learning, using social media data, be employed to forecast such events? In response, this study had as objective to develop a hybrid machine learning model to forecast protest and riot events in South Africa using Twitter data. Employing the CRISP-DM methodology, data was collected from Twitter for the period between 2019 and 2024, resulting in 18,487 curated tweets, with associated ground truth data extracted from the ACLED database. Using this data, a hybrid model combining Bidirectional LSTM (Bi-LSTM) networks with eXtreme Gradient Boosting (XGBoost) for classification and regression tasks was developed to forecast civil unrest in South Africa. Additionally, SHapley Additive exPlanations (SHAP) were used for model explainability. The proposed model outperformed the base model, achieving an R-squared value of 33% for protests and 23% for riots in regression, along with classification accuracies of 92% for protests and 86.2% for riots. SHAP results indicated that the key predictors of unrest included sentiment-related features, tweet engagement features, regional factors, the day of the week, public holidays, and the topics being discussed. This study demonstrates the value of a hybrid model in forecasting civil unrest events and identifies key features that stakeholders can use to target their efforts more precisely in addressing civil unrest, ensuring resources are allocated where they are needed most. The study concludes with a discussion of valuable insights for stakeholders on how to leverage social media data to predict and mitigate civil unrest.

This research note addresses a methodological gap in the study of large language models (LLMs) in social sciences: the absence of standardized data extraction procedures. While existing research has examined biases and the reliability of LLM-generated content, the establishment of transparent extraction protocols necessarily precedes substantive analysis. The paper introduces a replicable procedural framework for extracting structured political data from LLMs via API, designed to enhance transparency, accessibility, and reproducibility. Canadian federal and Quebec provincial politicians serve as an illustrative case to demonstrate the extraction methodology, encompassing prompt engineering, output processing, and error handling mechanisms. The procedure facilitates systematic data collection across multiple LLM versions, enabling inter-model comparisons while addressing extraction challenges such as response variability and malformed outputs. The contribution is primarily methodological—providing researchers with a foundational extraction protocol adaptable to diverse research contexts. This standardized approach constitutes an essential preliminary step for subsequent evaluation of LLM-generated content, establishing procedural clarity in this methodologically developing research domain.

Framing is among the most extensively used concepts in the field of communication science. The availability of digital data offers new possibilities for studying how specific aspects of social reality are made more salient in online communication, but also raises challenges related to the scaling of framing analysis and its adoption to new research areas (e.g. studying the impact of artificial intelligence-powered systems on the representation of societally relevant issues). To address these challenges, we introduce a transformer-based approach for generic news frame detection in Anglophone online content. While doing so, we discuss the composition of the training and test datasets, the model architecture, and the validation of the approach and reflect on the possibilities and limitations of the automated detection of generic news frames.

Systematic content analysis of messaging has been a staple method in the study of communication. While computer-assisted content analysis has been used in the field for three decades, advances in machine learning and crowd-based annotation combined with the ease of collecting volumes of text-based communication via social media have made the opportunities for classification of messages easier and faster. The greatest advancement yet might be in the form of general intelligence large language models (LLMs), which are ostensibly able to accurately and reliably classify messages by leveraging context to disambiguate meaning. It is unclear, however, how effective LLMs are in deploying the method of content analysis. In this study, we compare the classification of political candidate social media messages between trained annotators, crowd annotators, and large language models from Open AI accessed through the free Web (ChatGPT) and the paid API (GPT API) on five different categories of political communication commonly used in the literature. We find that crowd annotation generally had higher F1 scores than ChatGPT and an earlier version of the GPT API, although the newest version, GPT-4 API, demonstrated good performance as compared with the crowd and with ground truth data derived from trained student annotators. This study suggests the application of any LLM to an annotation task requires validation, and that freely available and older LLM models may not be effective for studying human communication.

Information disorder (IDO) presents a persistent challenge to society, necessitating innovative approaches to understanding its dynamics beyond just merely detecting it. This study introduces a theory-driven framework that integrates advanced natural language processing (NLP) with deep learning, utilizing the target-based emotion–stance analysis (TESA) approach to analyze emotion and stance dynamics within IDO content. Complementing TESA, interactive graph generation (IGG) is applied for scalable and interpretable qualitative analyses. Employing a mixed-methods approach, the study leverages TESA for target-centric emotion and stance analysis, evaluating target-based classifiers on both human-annotated and synthetic datasets. Additionally, the study explores synthetic data generation using generative AI to enrich the analysis, applying IGG to map complex data interactions. The study also found that integrating synthetic data developed from human annotations enhanced model performance, particularly for emotion classification tasks. Results demonstrate that IDO narratives significantly differ from non-IDO narratives, frequently leveraging negative emotions such as anger and disgust to manipulate public perception. TESA proved effective in capturing these nuanced variations, while IGG facilitated the triangulation of such findings via the scalable interpretation of emotional narratives, revealing that IDO content often amplifies polarizing and antagonistic perspectives. By combining TESA and IGG, this research emphasizes the importance of using NLP to extract and examine the emotional and stance nuances toward targets of interest within IDO context. This approach not only deepens theoretical insights into IDO’s persuasive mechanisms but also supports the development of practical tools for analyzing and managing the influence of IDO on public discourse.

Despite millions of hacked accounts fueling cybercrime, research on the hacking experience, particularly sociodemographic aspects, remains sparse. This study examines the experience of being hacked with a focus on gender disparities from the perspective of the third-level digital divide—socially constructed gaps of digital use outcomes even among users with similar digital access and skills. Analyzing 13,731 Twitter mentions of accounts being hacked, using topic modeling and classifying the gender of 12,586 users, we showed that women reported more experiences of being hacked across all types of online services except gaming. Women were more likely to experience negative consequences of being hacked, including reputational harm, money loss, and having personalized content modified. Gender differences were also found in coping strategies. Men were more likely to use active strategies like warning others, rebuilding accounts, and deducing hackers’ origins, while women were more likely to seek help from others to recover or report experiencing hacked accounts. The findings of this study imply the need for further research into the gendered experiences of being hacked from the third-level digital divide perspective, alongside the development of interventions to mitigate harm and empower users with diverse needs to cope with being hacked.

“Synthetic samples” generated by large language models (LLMs) have been argued to complement or replace traditional surveys, assuming their training data is grounded in human-generated data that potentially reflects attitudes and behaviors prevalent in the population. Initial US-based studies that have prompted LLMs to mimic survey respondents found that the responses match survey data. However, the relationship between the respective target population and LLM training data might affect the generalizability of such findings. In this paper, we critically evaluate the use of LLMs for public opinion research in a different context, by investigating whether LLMs can estimate vote choice in Germany. We generate a synthetic sample matching the 2017 German Longitudinal Election Study respondents and ask the LLM GPT-3.5 to predict each respondent’s vote choice. Comparing these predictions to the survey-based estimates on the aggregate and subgroup levels, we find that GPT-3.5 exhibits a bias towards the Green and Left parties. While the LLM predictions capture the tendencies of “typical” voters, they miss more complex factors of vote choice. By examining the LLM-based prediction of voting behavior in a non-English speaking context, our study contributes to research on the extent to which LLMs can be leveraged for studying public opinion. The findings point to disparities in opinion representation in LLMs and underscore the limitations in applying them for public opinion estimation.

Adolescents may perceive that social media exert influence on their beliefs, attitudes, and behaviors. Past research has found that frequent social media use and fear of missing out have related to risk behavior and poor mental health outcomes. Little research has been conducted on the perception of influence of social media by adolescents on mental health outcomes and risky behavior engagement. In this study, 304 adolescents (female = 210 and male = 94) completed an online questionnaire about their use of social media, perceptions of social media influence, fear of missing out, engagement in risky behavior, and depressive and anxiety symptoms. Age, perceptions of social media influence, and fear of missing out were significant predictors of engaging in risky behaviors. Age, being female, perceptions of social media influence, and fear of missing out predicted anxiety symptoms. Being female, perceptions of social media influence, and fear of missing out predicted depressive symptoms. For adolescents, the influence of social media on mental health outcomes and risky behaviors may be based on their perception of influence of social media and fear of missing out rather than just frequency of use.

Short-form videos have become a dominant form of social media globally. While short-video apps are popular among adolescents, their ease-of-use has also attracted a growing number of elderly users. However, this accessibility can lead to problematic use, resulting in physical and mental health issues for this demographic. Therefore, our research employed the technology acceptance model (TAM) to understand the problematic use of short-video apps (PUSVA) among elderly adults. 281 elderly adults completed a three-wave survey with a 1-month interval between waves. Results showed that both perceived utilitarian-usefulness and perceived hedonic-usefulness mediated the relationship between perceived ease-of-use and PUSVA, suggesting a double-edged sword effect of ease-to-use short-video apps. Moreover, perceived susceptibility moderated the relationship between perceived ease-of-use and perceived utilitarian-usefulness, but not between perceived ease-of-use and perceived hedonic-usefulness, suggesting a moderated mediation effect of perceived susceptibility on PUSVA. Specifically, elderly adults with low perceived susceptibility tended to report higher perceived utilitarian-usefulness for easy-to-use applications, while no relationship between perceived ease-of-use and perceived utilitarian-usefulness was observed among those with high perceived susceptibility. Our findings highlight the double-edged sword effect of user-friendly short-video apps and offer valuable insights for developing interventions to mitigate problematic use among elderly adults.

Prior research has largely documented the overall mobilizing effects of social media news consumption and political discussion linked to citizens’ political participatory behaviors. However, limited empirical research has considered the informational and communicative effects to be contingent upon different social media platforms. Therefore, this study advances distinct theoretical affordances and effects of social media news use on online (by using online versions of legacy media outlets, blogs, and news apps) and social media political participation. Taking advantage of US comparative panel data, ordinary least squares (OLS) causal autoregressive regressions and panel autoregressive structural equation model tests cast a much-needed light on the diverse effects of Facebook, X, Snapchat, WhatsApp, Instagram, YouTube, and Reddit use for news over both political discussions with weak and strong ties, and political participation online and in social media. Moreover, results from two-step algorithmic cluster analysis clarify how these social media platforms generate different information and political behavior clusters of citizens, which also provide a comparative view of how social media platforms differently contribute to people’s public and political life in US democracy.

Theoretical frameworks Resource-Based View (RBV) and competitive advantages have served as conceptual foundations for investigating the role of Google Maps in business success. This research has two key findings: First, an analysis of a dataset obtained by scraping local business information from Google Maps ( N = 9,445) shows that minority-owned businesses were less likely to be claimed on Google Maps and received fewer consumer review comments compared to their non-minority counterparts. Second, a comparison of Google Maps data collected before the outbreak of the COVID-19 pandemic in 2019 and follow-up data gathered in 2022 reveals higher survival rates among businesses that were claimed, utilized business attributes, or had more reviews. Together, these findings suggest that a stronger presence on Google Maps contributed to competitive advantages and business survival during the pandemic. This underscores the importance of Google Maps presence for the survival of businesses during a crisis.

Media bias has long been a subject of scholarly interest due to its potential to shape public perceptions and behaviors. This systematic review leverages advances in natural language processing (NLP) to explore automated methods to detect media bias, addressing five core questions: it examines the definitions and operationalization of media bias, explores the NLP tasks addressed for its detection, the technologies used, and their respective outcomes and applied findings. This review also examines the practical applications of these methodologies and assesses the patterns, implications, and limitations associated with using artificial intelligence for media bias detection. Analyzing peer-reviewed articles from 2019 to 2023, the review initially identified 519 articles, which ultimately included 28 relevant ones. Significant heterogeneity is observed in bias definitions, affecting the analysis and detection approaches. The review highlights the predominant use of some methods and identifies challenges such as inconsistencies in problem definition, outcome measurement, and comparative method evaluation. Regardless of the conceptualizations of bias and the methods used, studies consistently identify bias in media outlets. Thus, studying media bias remains necessary for raising awareness and detection, and NLP methods are significant allies in this endeavor. This research aims to consolidate the foundations of recent advances in NLP for bias detection, encouraging researchers to focus on developing transparent, task-specific tools and work toward a consensus on a technical definition of bias and standardized metrics for its evaluation.

This study examines electoral forecasting in volatile party systems, focusing on factors contributing to deviations between poll predictions and actual election outcomes. Using Italy as a case study, it identifies biases in polling data and proposes a method to enhance estimator accuracy in a context of stable institutions and volatile electoral dynamics. Data from three Italian general elections are analyzed to evaluate discrepancies between pre-electoral polls and results, assessing key factors such as timing of data collection, survey methodology, sample size, and party system fragmentation. Employing a Bayesian inference process via a Markov chain Monte Carlo (MCMC) adaptive Metropolis-Hastings (MH) algorithm, the study demonstrates that pre-electoral estimates can be significantly improved using the Two-Stage Model (TSM). By consistently outperforming traditional poll predictions, the TSM offers a robust framework for addressing polling biases. These findings advance political forecasting by improving accuracy in both consolidated democracies and volatile electoral contexts, while emphasizing the need for future research on dynamic polling methods and fundamentals-based models.

Large language models show promising capability in some qualitative content analysis tasks; however, research reporting their performance in identifying initial codes that underpin subsequent analysis is scarce. This paper explores the suitability of GPT-4 to assist in building a codebook for a discourse network analysis (DNA) of a recent alcohol policy reform. DNA is a codebook-driven approach to identifying groupings of actors who use similar policy framings. The paper uses GPT-4 to identify initial codes (‘concepts’) and related quotes in 108 news articles and interviews. The results produced by GPT-4 are compared to a codebook prepared by researchers. GPT-4 identified over two-thirds of the concepts found by the researchers, and it was highly accurate in screening out a large volume of irrelevant media items. However, GPT-4 also provided many irrelevant concepts that required researcher review and removal. The discussion reflects on the implications for using GPT-4 in codebook preparation for DNA and other situations, including the need for human involvement and sample testing to understand its strengths and limitations, which may limit efficiency gains.

Understanding reporting behavior in questionnaires is a key issue in enhancing cross-national data comparability and policy decisions. Computers help improve the analysis of careless or insufficient effort (C/IE) responding by logging response times and other response behavior, ensuring data quality. We introduce a response-time based approach, built on an analysis of the relationship between a survey item and a related external variable, to cross-national research. Using PISA 2015 data from 58 countries/economies, we analyze patterns of correlations between the enjoyment of science and science test scores across response time. We focus on C/IE responding towards the beginning of the response time spectrum. Results indicate rather diligent responding in Eastern Asia and a part of Northern Europe. Yet in other regions (e.g., part of Latin America and the Caribbean, and Eastern Europe), C/IE responding might be distorting the data. We provide other researchers with information regarding when and to what extent C/IE responding can occur across countries. We enhance the understanding of heterogeneity in reporting behavior across countries.

Research indicates that prejudice has been growing in America. Citizens feel increasingly threatened by immigrants, and hate crimes against immigrant groups have risen. Declining interpersonal contact has also made it more difficult to address prejudice directly. This study examines whether nonpolitical social media groups can foster connections that reduce prejudice. These groups allow users to connect on the basis of shared interests, enabling diverse individuals to form close relationships which may improve attitudes toward immigrants. Using a national web survey matched to U.S. Census percentages for sex, race, ethnicity, age, and region of residence ( N = 1500), along with a two-wave panel conducted over six weeks ( N = 752), results indicate that blatant prejudice is more prevalent than subtle prejudice. Respondents were more likely to feel threatened by immigrants than to withhold positive emotions from them. As a remedy, social connectedness in nonpolitical groups was associated with diminished blatant prejudice and lower levels of global prejudice, a measure that includes both subtle and blatant components. Findings suggest that feeling connected with different people remotely can improve attitudes toward racial and ethnic diversity, helping individuals feel less threatened by immigrants and less prejudiced overall.

As the application of artificial intelligence in various domains and sectors grows, politics—especially political communication—is no exception. However, academic considerations on the topic remain limited, partly due to its novelty. To contribute to the ongoing discussions at the intersection of AI and political campaigns, this research report presents the development and use of an AI chatbot employed by an Italian candidate during the 2024 European Parliament elections. The aim of this work is to engage with the technical aspects of the tool’s development and implementation by outlining the challenges and strategies involved in creating an AI chatbot that supports a political campaign using OpenAI APIs. Furthermore, this report offers reflections on the role of AI in politics and communication, focusing on the concepts of intermediation and participation, also addressing issues of compliance and trustworthiness of these new AI tools.

The intersection of politics and religion in India has gained significant scholarly attention, particularly since the Bharatiya Janata Party’s (BJP) rise to power in 2014. The increasing impact of social media on Indian politics has intensified this concern. However, it is yet to be fully explored how social media was used for religiopolitical purposes during the Indian election in 2024. We computationally analyzed 3082 Facebook posts using BERTopic, word embedding, and cluster analysis to understand how politicians, political candidates, political organizations, and political parties intertwined religion and politics during the 2024 Lok Sabha election. We identified the presence of religiopolitical propaganda, primarily aimed at reviving and recreating Hindu nationalist history and targeting religious minorities, mainly Muslims. The major topics of the posts included ideological legacy, political landscape, party and leadership, celebrations, crime and justice, local politics and governance, politicized demographic trends, public engagements, spiritual and philosophical themes, and the misrepresented reservation issue. The interconnectedness of these issues suggests that the BJP and its allies concentrated on religious matters, from Hindu–Muslim debates to reservations for Muslims and the inauguration of Hindu temples. Data from non-political entities, such as influencers, as well as cross-platform analysis from Twitter and YouTube, can extend and enrich these insights.

Understanding the dynamics and duration of trending topics on digital platforms has been deemed a crucial issue of research in computer-mediated communication. Employing event history analysis (EHA), this study attempts to examine the antecedents of the lifespan and evolution of trending topics on Sina Weibo, one of the most prevalent social media platforms in China. Based on a collection of 2,386 Sina Weibo trending topics that emerged between January 1 and January 31, 2022, the study explores how factors of trending topics, especially their semantic and textual features, significantly influence topic persistence. Our findings indicate that the median survival time of Weibo trending topic was 6.28 hours, with an average of 8.29 hours. In addition, exogenous non-viral topics, driven by external events or media coverage, tend to remain on the trending list longer than other topics. Furthermore, political topics tend to have a relatively longer period of survival duration when compared to social events, life records and moods, and fashion and entertainment topics. Lastly, actionable and opinion-oriented topics tend to have shorter longevity compared to informational and emotional topics. By quantifying the factors that affect trending topic duration, the study offers a novel theoretical perspective on the role of political drivers and external influences in shaping collective and connective digital discourses. The findings contribute to the broader field of digital platform communication, particularly in public opinion management and content governance on social media.

Social media has become an integral part of daily life, shaping behaviors, self-perception, and emotional well-being. However, its addictive use raises concerns about its potential to aggravate psychological challenges, particularly in the context of societal expectations of masculinity. The current report presents a study exploring the pathways through which social media addiction contributes to masculine depression, specifically examining the roles of physical appearance comparison, self-esteem, and emotional control among men. By investigating these relationships, it aims to provide insights into the psychological consequences of social media addiction for men. Structured questionnaires were administered to 849 Israeli men aged 18 and older. Employing a moderated sequential mediation model with social media addiction as the independent variable, physical appearance comparison and self-esteem as mediators, and masculine depression as the dependent variable, this study also investigates emotional control as a moderator in the associations between social media addiction, physical appearance comparison, self-esteem, and masculine depression. The analysis, conducted using model 89 PROCESS v4.2 macro, reveals that conforming to the masculine norm of emotional control intensifies men’s vulnerability to distress resulting from maladaptive behaviors such as social media addiction, which can lead to masculine depression. Furthermore, addiction to social media can trigger masculine depression via psychosocial factors such as physical appearance comparison and low self-esteem, which have yet to be explored in the context of masculine depression. These findings underscore the importance of targeted interventions that address the societal pressures of masculinity and the psychological repercussions of excessive social media use among men. They also emphasize the necessity of raising awareness about these issues among both the public and therapists.

The proliferation of misinformation in the digital age has emerged as a pervasive and pressing challenge, threatening the integrity of information dissemination across online platforms. In response to this growing concern, this survey paper offers a comprehensive analysis of the landscape of misinformation detection methodologies. Our survey delves into the intricacies of model architectures, feature engineering, and data sources, providing insights into the strengths and limitations of each approach. Despite significant advancements in misinformation detection, this survey identifies persistent challenges. The paper accentuates the need for adaptive models that can effectively tackle rapidly evolving events, such as the COVID-19 pandemic. Language adaptability remains another substantial frontier, particularly in the context of low-resource languages like Chinese. Furthermore, it draws attention to the dearth of balanced, multilingual datasets, emphasizing their significance for robust model training and assessment. By addressing emerging challenges and offering a comprehensive view, our paper enriches the understanding of deep learning techniques in misinformation detection.

Despite being influential spaces for disseminating information, social media platforms often contribute to the perpetuation of harmful stereotypes and misconceptions surrounding depression. This study investigates the relationship between exposure to stigmatizing depression on social media and help avoidance among young adults, while examining the mediating roles of depression knowledge and stigmatizing attitudes. A sample of 428 Chinese young adults aged between 18 and 35 responded to the anonymous questionnaires. Results indicate a positive association between exposure to stigmatizing information on social media and help avoidance. Furthermore, depression knowledge and stigmatizing attitudes were found to mediate this relationship , highlighting the cognitive mechanisms underlying the impact of social media on mental health attitudes and behaviors. The findings underscore the importance of addressing stigmatizing content on social media platforms and promoting accurate depression knowledge among young adults to mitigate help avoidance tendencies. Implications and limitations are discussed.

This paper describes the creation of a novel dataset on ministerial turnover and resignation calls in 12 presidential cabinets in Latin America from the mid-1970s to the early 2020s. The indicators on resignation calls and reallocations of cabinet members are entirely novel. Both constitute a relevant empirical contribution not only to the study of political dynamics in presidential systems and cabinet politics but also to public opinion and public policy topics. We focus on the creation of the dataset using optical recognition algorithms on press report archives together with machine learning models. The models permitted the training of ensemble semi-supervised classifiers over a period of almost 50 years. Subsequently, we provide a number of measurement validity checks to cross-validate the dataset by comparing it with similar existing data and an exploratory analysis.