As large language model (LLM) agents become routine participants in public online conversation, understanding how their replies reshape human-to-human interaction structure is a central challenge for social computing and platform governance. Yet most post-deployment evaluations focus on toxicity or engagement volume, leaving thread-level conversational structure—a dimension critical to deliberative quality, social capital formation, and equality of voice—largely unmeasured. We address this gap through a large-scale structural evaluation of @CommentR , a production LLM agent serving millions of users on Weibo, China’s leading microblogging platform. Modeling each thread as a directed reply graph, we match over 216,000 real post-deployment threads on strictly pre-anchor covariates and estimate effects on human-only conversational structure using a doubly robust estimator. Because agent replies often arrive before any human-to-human interaction is observed, we introduce lifecycle-aware estimands that distinguish early-stage formation effects from mature-thread rewiring effects. Under conditional ignorability, agent replies reduce reciprocity, increase branching, and reduce geographic homophily in early-stage threads, while degree-corrected bridging remains unchanged—consistent with a hub-and-spoke shift from dialogic exchange toward one-off commenting around a focal reply. In mature threads, eligibility for the incumbent-rewiring analysis is itself reduced by the agent—sustained incumbent exchange becomes less likely—so we report the mature-thread rewiring estimates as bounded rather than point-identified and treat the formation regime as the one in which our structural evidence is secure. The magnitude of the reshaping depends on how the agent answers: more comprehensive, factual replies produce a larger focal shift. Reply-target analysis confirms that a substantial share of human comments redirect toward the agent, reducing lateral human-to-human exchange. These findings demonstrate that public AI agents can reshape not only what people say but how people talk to each other , with direct implications for platform governance, conversational agent design, and the structural monitoring of AI-mediated public discourse.
The ubiquity of generative AI (GenAI) requires social computing scholarship to examine how such extensive AI mediation may (re)shape the emotional and analytic qualities of human communication, both of which are central to audience engagement, trust, and information integrity. This study conducts a large-scale evaluation of 11 custom and fine-tuned LLMs, including DeepSeek, Llama (Meta), Mistral, and Gemma (Google), by asking them to rewrite the complete set of content, published by a local news outlet over a 12-year span (2011–2023). Using transformer-based sentiment models (i.e., RoBERTa fine-tuned on GoEmotions) and lexicon-based measures (i.e., LIWC, NRCLex), we compare the emotional expressiveness and analytic style of human-written versus AI-adapted content across short-form (titles/headlines) and long-form communication (article bodies/content). Analyzing 3,623 original news articles, nearly 40,000 primary-prompt AI-generated rewrites, and an additional generic-prompt ablation set, we estimate linear mixed-effects models that account for article-level clustering and control for model and prompt heterogeneity, communication form, verbosity, and readability. Results show that, under the corpus and prompting conditions examined here, LLMs tend to intensify emotional expression relative to human-written texts. Meanwhile, LLM rewrites tend to preserve, and in many cases increase, linguistic markers associated with analytic style relative to human benchmarks. Interestingly, even under the simpler generic prompt, LLMs in our sample display a shift toward more positive emotional categories, including joy, excitement, admiration, optimism, and gratitude, whereas human writers in this corpus display a broader affective range and more nuanced expression. To assess risks of clickbait-like headline framing, we fine-tune a BERT model for clickbait detection and examine linguistic cues in titles and headlines, yet we find minimal evidence that LLMs’ emotional amplification devolves into sensational and hyperbolic framing. Lastly, to enhance validity, we conduct a blinded human evaluation study on a randomly sampled subset of the corpus. Across 17 small experiments, human judges ( n = 170) independently evaluated analytic style, reasoning depth, and clickbait characteristics. These human evaluations provide convergent evidence that AI-generated texts were perceived as more analytically structured and slightly higher in perceived reasoning depth, while AI-generated headlines were also rated as less clickbait-oriented. Overall, our findings suggest that GenAI and LLMs have the potential to enhance the affective intensity of communication without necessarily compromising its informational quality. However, as human communication becomes increasingly and inevitably AI-mediated, current LLMs may shift the affective register, narrow the emotional spectrum, and introduce tonal biases in public communication that could subtly influence how information is disseminated, and how language is traditionally interpreted.
As large language models (LLMs) become deeply embedded in daily life, the urgent need for safer moderation systems that distinguish between naive and harmful requests while upholding appropriate censorship boundaries has never been greater. While existing LLMs can detect dangerous or unsafe content, they often struggle with nuanced cases such as implicit offensiveness, subtle gender and racial biases, and jailbreak prompts, due to the subjective and context-dependent nature of these issues. Furthermore, their heavy reliance on training data can reinforce societal biases, resulting in inconsistent and ethically problematic outputs. To address these challenges, we introduce GuardEval, a unified multi-perspective benchmark dataset designed for both training and evaluation, containing 106 fine-grained categories spanning human emotions, offensive and hateful language, gender and racial bias, and broader safety concerns. We also present GemmaGuard (GGuard), a Quantized Low-Rank Adaptation (QLoRA), fine-tuned version of Gemma3-12B trained on GuardEval, to assess content moderation with fine-grained labels. Our evaluation shows that GGuard achieves a macro F1 score of 0.832, substantially outperforming leading moderation models, including OpenAI Moderator (0.64) and Llama Guard (0.61). We show that multi-perspective, human-centered safety benchmarks are critical for mitigating inconsistent moderation decisions. GuardEval and GGuard together demonstrate that diverse, representative data materially improve safety, and adversarial robustness on complex, borderline cases.
The growing volume of socially generated content–ranging from online discussions to news articles–poses increasing challenges for collective understanding and information synthesis. Multi-document summarization offers a means to distill key ideas from such large and redundant text collections, yet existing large language models struggle to integrate dispersed context and maintain factual consistency. This article presents a modular summarization framework that integrates dynamic spectral clustering with structured abstractive rewriting to improve both interpretability and coherence. Our method first partitions related documents into semantically coherent clusters using a dynamic spectral algorithm inspired by community detection, then produces localized summary drafts with a lightweight sequence-to-sequence model, and finally refines these drafts through an abstractive rewriting stage. This structured design enables controllable and transparent summarization across heterogeneous sources. Experiments on benchmark datasets demonstrate that the proposed approach achieves competitive results relative to recent large-model systems while offering clearer interpretability and efficiency advantages for social computing applications.
Technological advancements have greatly impacted labor market dynamics, leaving a psychological impact on workers. Although some studies have explored such labor market changes and their effects on workers, they are limited to self-reported data, such as surveys and questionnaires. In this article, we propose a new approach for identifying information workers’ challenges and their impact on workers’ emotional well-being using large-scale, inexpensive, and near-real-time online social netwok data. While the research is situated within the broader Future of Work context shaped by technological change, the data collected and analyzed focus specifically on IT-related skills and topics discussed on Reddit, rather than on artificial intelligence itself. Analyzing over 700,000 Reddit posts related to IT occupations, we identify major labor market topics, including education and skill development, job search, and employment concerns, and examine how workers’ emotional expressions vary across gender and age groups. Our findings reveal systematic differences in how workers from different demographic groups disclose their needs and emotional states online. This work demonstrates the value of social media data as a complementary lens for understanding workers’ challenges in the knowledge economy. It highlights the potential of online communities as vehicles for peer support and well-being interventions.
Reproducibility of crowdsourced labeling results across platforms remains a challenge, as differences in worker ability distributions often lead to inconsistent accuracy. This article addresses this issue by proposing an ability-based approach to improve reproducibility. We hypothesize that matching worker ability distributions across platforms enhances result consistency. Using Latent Rank theory, we model worker abilities from original experiments and create a benchmark test to select participants for replication studies. Our findings show that using worker sets with similar ability distributions significantly improves reproducibility. However, requiring all workers to answer all tasks is impractical. To address this, we develop a collaborative filtering method that reduces the number of required workers to 3–5 per task, enabling reproducibility without additional cost. This approach enables more reliable crowdsourcing for scientific research and AI development.
The digitalisation of the reproductive body has engaged myriads of cutting-edge technologies to support people in understanding and managing their intimate health. Generally understood as female technologies (FemTech), these products and systems collect a wide range of intimate data, predominantly related to sexual and reproductive health. This sensitive data is then processed, transferred, saved, and shared with other parties. The data-hungry nature of this industry and the lack of proper safeguarding mechanisms, standards, and regulations for this vulnerable data can lead to complex harms or diminished user agency. In this article, we adopted mixed methods, including an online survey and a Story Completion Method (SCM), to explore user understanding of the security and privacy (SP) of these technologies. We conducted online user studies in two countries: the UK and Iran, in English and Farsi (Persian), respectively. We recruited 102 UK and 130 Iranian participants for our surveys, and 17 UK and 34 Iranian participants for our SCM studies. Our findings show that while users can speculate on the range of harms and risks associated with these technologies, they are not necessarily presented and equipped with the technical skills to protect themselves. We identified a larger gap in the Iranian group compared to the UK group, which can be attributed to factors such as differences in data protection regulations between the two countries. This comparative study highlights how disparate data protection frameworks and unique cultural contexts can shape user vulnerability to FemTech risks. We discuss how participatory threat modelling and Security and Privacy by Design are critical to protecting users in these sensitive systems.
Research has consistently shown that websites’ privacy policies fail to inform users, prompting to explore more condensed, visual solutions such as icons and privacy labels. However, we miss efficient and scalable systems to generate such privacy labels and conclusive evidence on their impact on people’s privacy protection behavior. This article fills these gaps: we present the Visual Privacy browser extension that automatically computes a privacy rating and visualizes it in a privacy label. We studied the impact of those labels on privacy protection behavior in an online experiment ( N =825). Our results show that the privacy labels shown by Visual Privacy significantly impact a website’s perceived risks and benefits, and that they have an indirect effect on privacy protection behavior. Study participants showed high awareness and acceptance of privacy labels, with 79% of participants recalling the rating shown in a label and over 75% reporting that they would like them to become mandatory.
The wisdom of crowds – the finding that aggregating judgments across individuals often outperforms the best individual – has been extensively studied with human forecasters. Whether the same phenomenon emerges when the “crowd” consists of large language models (LLMs) is an open question with both theoretical and practical implications. We elicited probability estimates from 15 LLMs on 254 binary prediction market questions and evaluated classical and learned aggregation methods. Learned aggregators – a multilayer perceptron and a logistic regression – outperformed all individual models and classical methods. The logistic regression was found to match the neural network, suggesting that the benefit of learned aggregation derives from learning a linear combination of diverse model outputs rather than from nonlinear interactions. Symbolic regression applied to the neural network's learned mapping recovered a pure model-disagreement signal as the lowest-complexity useful formula on the Pareto frontier, further supporting this interpretation. Training cutoff contamination proved a pervasive confound: the apparent capability gap between frontier cloud models and smaller local models collapsed from 35.8
Menopause affects half of the world’s population directly and everyone indirectly. As the world’s population is ageing, more healthcare services are becoming digitised. People experiencing menopause seek online and digital help for various reasons. Menopause education, awareness, and research – including the usage and risks of menopause tech and data, are limited. In this work, we designed and conducted a large-scale online survey with 310 UK participants to investigate the experiences of women with menopause, the use of technology, and security and privacy perceptions and expectations regarding their data. We identify several factors influencing the use of menopause tech, including perceived benefits, community support, workplace, and healthcare. Users express concerns about misinformation and current data sharing practices associated with these technologies. Conditional consent for data access and sharing is prevalent, depending on the recipient and user control options. This research provides crucial empirical data and analysis on menopause tech use, security, and privacy while identifying user concerns and informing secure user-centric design and policy.
Fake news is false or misleading information shared as real news, often meant to trick or influence public opinion. Its fast spread through social media causes confusion, divides communities, and reduces trust in reliable news sources. While much research has been done on fake news detection in English, there is a lack of tools designed specifically for Bangla using a combined linguistic and psychological approach, despite the rise of fake news in Bangla-speaking regions. Our research aims to bridge this gap by developing tailored tools that effectively address the nuances of fake news in the Bangla language. By implementing a robust detection model, we contribute to a deeper understanding of misinformation in this context. In this work, we present a psycholinguistic method that combines linguistic and psychological features of the Bangla language to detect fake news. We created a dataset of 5,428 Bangla news articles, categorized as real or fake, capturing diverse topics and styles. This research highlights the importance of developing language-specific tools to combat misinformation. It also offers a valuable resource for future studies, helping to promote a more informed and responsible media environment in Bangla-speaking communities. Our hybrid model, which integrates Convolutional Neural Networks (CNN) and Multilayer Perceptron (MLP), effectively captures these features, achieving an accuracy of 96.32%. These results demonstrate the strong potential of our approach in improving the detection of fake news in Bangla.
The increasing relevance of valued networks, which account for both the presence and intensity of relations, spans various disciplines, from life sciences to engineering and social sciences. However, comparing such networks remains a methodological challenge due to the lack of effective measures for quantifying differences. This paper introduces the ξ -distance, a novel metric designed to measure dissimilarity between valued networks. The proposed approach extends current graph comparison techniques by integrating both topological and relational intensity features into a single framework. Unlike existing measures, ξ -distance efficiently handles directed and weighted networks, employing a probabilistic framework based on vertex distance sequences (VDS). This approach is demonstrated through extensive simulations and real-world applications, including international trade and social interaction networks, where ξ reliably differentiates networks based on structural and relational characteristics. Moreover, the ξ -distance is computationally efficient, with an implementation that combines R and C++, ensuring accessibility and high performance. This metric not only addresses the limitations of previous methods but also opens new avenues for analysing multiplex, dynamic, and complex network systems, offering significant improvements in network comparison tasks across multiple domains.
Online platforms urgently need better ways of managing inciting speech that will not unreasonably limit lawful political speech. To address this, we use discourse on the minimally-moderated platform Gab surrounding the Charlottesville Unite The Right Rally in 2017. We first adapt the Dangerous Speech Paradigm to a three-tiered framework with messages classified as either allowable , harmful , or inciting . The bisection of the previous ”dangerous” category into ”harmful” and ”inciting” enables more targeted classification. Using a customized GPT-4o-mini large language model (LLM) framework that utilizes Chain-of-Thought prompting (CoT) and in-context examples, we classify over 3.5 million posts under the three-tiered framework. To validate the compatibility of our framework with diffusion methods which may be used in future work, we then build a directed repost graph to model the spread of harmful and inciting speech through the use of a DeGroot diffusion model. We find support for LLM classifiers to assist human content moderators in identifying incitement according to our framework. With this assistance, intervention methods can be more effectively targeted, better protecting the public while preserving freedom of speech.
In the era of digital governance, social media platforms have become critical arenas for citizen-state interaction and policy feedback. However, governments often struggle to translate massive, unstructured public discourse into actionable intelligence, frequently relying on aggregated metrics that obscure critical regional nuances. This study introduces a spatially-aware computational framework to bridge this gap. Integrating Large Language Model (LLM) based few-shot learning with SHAP (SHapley Additive exPlanations) and GIS analysis, we examine the public reception of China’s 2025 national childcare subsidy policy using a dataset of 352,448 comments. Our analysis reveals a stark divergence between information-seeking behaviors (high volume) and substantive legitimacy debates (high engagement). Crucially, we demonstrate that the socioeconomic drivers of this discourse—such as gender ratios and educational attainment—are highly spatially heterogeneous, necessitating distinct governance responses across different regions. This research contributes to the field of government information systems by validating a novel, replicable methodology for precision policy analytics, offering policymakers a tool to move beyond national averages and achieve more responsive, context-specific governance.
African American teens make up a large demographic on social media platforms - more specifically TikTok. Prior research in HCI and social computing indicates that social media can play a large role in contemporary teen identity formation through both self-expression and self-perception. We build on this research as well as Black girlhood studies by exploring the identity-relevant and culturally-relevant experiences Black teens have through a digital ethnography of TikTok. We study this through the lens of Black prom culture, a historical African American tradition, seen through the #HoodProm and #BlackProm hashtags. We collected and analyzed 219 TikTok videos, observing how Black teens showcase joyful narratives and culturally preserve African American culture on the internet.
Flagging mechanisms on social media platforms allow users to report inappropriate posts/accounts for review by content moderators. These reports are pivotal to platforms' efforts toward regulating norm violations. This paper examines how platforms' design choices in implementing flagging mechanisms influence flaggers' perceptions of content moderation. We conducted a survey experiment asking US respondents (N=2,936) to flag inappropriate posts using one of 54 randomly assigned flagging implementations. After flagging, participants rated their fairness perceptions of the flag submission process along the dimensions of consistency, transparency, and voice (agency). We found that participants perceived greater transparency when flagging interfaces included community guidelines and greater voice when they incorporated a text box for open-ended feedback. Our qualitative analysis highlights user needs for improved accessibility, educational support for reporting, and protections against false flags. We offer design recommendations for building fairer flagging systems without exacerbating the cognitive burden of submitting flags.
Designers’ use of deceptive and manipulative design practices have become increasingly ubiquitous, impacting users’ ability to make choices that respect their agency and autonomy. These practices have been popularly defined through the term “dark patterns” which has gained attention from designers, privacy scholars, and more recently, legal scholars and regulators. The increased interest in the concept of dark patterns across a range of practitioners and users motivated us to study the evolution of the concept and highlight the future trajectory of conversations around dark patterns and similar activist movements that utilize social computing platforms. In this paper, we examine the history and evolution of the Twitter discourse around #darkpatterns from its inception in June 2010 until April 2021, using a combination of quantitative and qualitative methods to describe how this discourse has changed over time. We frame these conversations through a new concept of socio-technical activism, whereby participants unite in order to identify and fight back against problematic technology and design practices. We discuss the potential future trajectories of this discourse and opportunities for further social computing scholarship at the intersection of design, policy, and online activism.
Adversary emulation is commonly used to test cyber-defense performance against known threats to organizations. However, many adversary emulation methods often rely on automated planning and underplay the role of human cognition. Consequently, defenders are often underprepared for human attackers who can think creatively and adapt their strategies. In this paper, we propose the design of adversarial cognitive agents that are dynamic, adaptable, and able to learn from experience. These cognitive agents are built based on the theoretical principles of Instance-Based Learning Theory (IBLT) of experiential choice in dynamic tasks, making them more challenging than strategically optimal adversaries for human defenders. Our research offers three main contributions. First, in a simulation experiment, we demonstrate how IBL attacker agents can learn from experience and become as efficient as optimal strategic algorithms against a strategic defender. In a second simulation experiment, the IBL attackers are pitted against an IBL defender, showing that the IBL attacker can be a more challenging adversary for the IBL defender, while the IBL defender can learn to counter carefully crafted optimal attack strategies. To test these observations, we conducted a third experiment, where humans played the role of defenders against both strategic and IBL attackers in an interactive task. The results confirm the predictions of the second simulation experiment: a cognitive attackers are more challenging for human defenders than strategic attackers. These insights contribute to informing future adversary emulation efforts and training of cyber defenders.
Microblogging platforms have been increasingly used by the public in crisis situations, enabling more participatory crisis communication between the official response channels and the affected community. However, the sheer volume of crisis-related messages on social media can make it challenging for officials to find pertinent information and understand the public’s perception of evolving risks. To address this issue, crisis informatics researchers have proposed a variety of technological solutions, but there has been limited examination of the cognitive and perceptual processes and subsequent responses of the affected population. Yet, this information is critical for the crisis response officials to gauge public’s understanding of the event, their perception of event-related risk, and perception of incident response and recovery efforts, in turn enabling the officials to craft crisis communication messaging more effectively. Taking cues from the Protective Action Decision Model, we conceptualize a metric —resonance+ — that prioritizes the cognitive and perceptual processes of the affected population, quantifying shifts in collective attention and information exposure for each tweet. Based on resonance+, we develop a principled, scalable pipeline that recommends content relating to people’s cognitive and perceptual processes. Our results suggest that resonance+ is generalizable across different types of natural hazards. We have also demonstrated its applicability for near-real time scenarios. According to the feedback from the target users, the local public information officers (PIOs) in emergency management, the messages recommended by our pipeline are useful in their tasks of understanding public perception and finding hopeful narratives, potentially leading to more effective crisis communications.
