Effective human-AI collaboration, especially in failure scenarios, requires systems that function as active partners rather than static tools. This research addresses this requirement by introducing a tiered multi-agent architecture grounded in Human-Centered eXplainable AI principles. The architecture consists of three novel design artifacts: tiered reasoning that adapts explanation depth to failure severity, a traceable memory bus for auditability, and flexible reasoning tools for enhanced adaptability. These designs enable agents to not only perform evidence-based diagnoses of performance gaps but devise recovery strategies and propose actionable improvement plans as well. We empirically evaluate this architecture on an aspect term extraction task using hybrid methods that combine performance comparisons against state-of-the-art baselines, human expert user studies, and multi-role user simulations. The results demonstrate that our architecture significantly enhances both task performance and failure recovery diagnosis and plans. We then validate the generalizability of our architecture with a second task of comparable complexity. This research contributes an adaptable and generalizable framework and foundational artifacts for trustworthy AI teammates.
Chatbots are increasingly used to support emotional well-being. This study examines how formal communication style and chatbot identity label influence users’ perceptions of emotional support. We developed Airbud, a generative AI chatbot presented under different identity labels (Generative AI vs. traditional chatbot) and collected data from 450 international students using formal and informal communication styles. Using structural equation modeling, the results show that formal communication style enhances perceived tie strength, while perceived warmth strengthens its effect on social presence, perceived tie strength, and emotional relief. Interaction satisfaction shapes perceived chatbot support for loneliness, sadness, and anxiety. Perceived chatbot support refers to users’ perceived support efficacy rather than actual clinical improvement or measurable emotional change. Multi-group analysis shows that formal communication style is less strongly related to perceived judgment under the GenAI label. These findings highlight the importance of communication style, perceived warmth, and identity labels in chatbots for international students.
Cross-domain R D collaboration in urban mobility is stalled by disciplinary silos that existing foresight methods cannot bridge. This study proposes a multi-agent large language model (LLM) framework that converts patent-based convergence signals into actionable industry–academia R D proposals. Using 26,399 WIPO patent applications (2000–2024), the system identifies high-potential technology pairs via Bayesian long short-term memory (LSTM) signal detection, builds expert personas from inventor records, retrieves academic collaborators through ArXiv queries, and produces structured R D roadmaps through LangGraph-orchestrated three-phase dialogues. A Two-Track evaluation separates an upper-bound capability demonstration (Track A) from primary validity evidence (Track B), which employs six non-generative discriminative scorers to substantially reduce evaluator–generator circularity (pipeline advantage: Δ = + 0.275, d = 2.227). Expert evaluation confirms strong inter-rater reliability (intraclass correlation = 0.918, n = 24), though a calibration gap indicates automated scores are systematically higher than human ratings. Limitations include short-window signal detection, proxy-level automated scoring, and unconfirmed real-world feasibility.
Social media platforms shape political discourse, yet their emotional and multimedia affordances also create fertile ground for misinformation. This study examines how emotional resonance and multimedia cues drive engagement and within-platform virality in Reddit’s largest political communities, offering implications for theory and governance in the Generative AI era. Using 17,548 posts and 1.5 million comments, we apply deep learning models (BERT, DistilRoBERTa) to capture fine-grained affective expressions. Results show that negative emotions—especially anger, disgust, and fear—are strongly associated with engagement and within-platform virality, while positive emotions have weaker effects. Multimedia presence significantly increases emotional intensity and reach, highlighting the importance of peripheral multimedia cues in emotional engagement and amplification. Additional associations involving moderator status and account-level authentication characteristics suggest that governance-related features may also be linked to lower emotional amplification. These findings clarify organic amplification dynamics that become especially consequential in contemporary social media environments.
Competition among software firms has given rise to theft of intellectual property and numerous patent infringement lawsuits. A potential key threat is the unauthorized use of code, ideas, or program components by competitors. If former employees or outsourcing partners steal their employers’ intellectual capital, organizations’ software product development may be compromised as a result. This unauthorized use of previous employers’ intellectual capital is a violation of non-disclosure agreements (NDA) by software developers, but IS research has offered limited explanation of this phenomenon. As a first step in addressing this gap, we present a model explaining NDA violations based on the social learning theory (SLT) of criminology. The new insight derived from SLT focuses on the motivations for violation, rather than on how organizations can deter such violations, which is the dominant theoretical lens in IS security to address information policy violations. We argue that, owing to the inconspicuous nature of NDA violations, deterrence (theory) alone is insufficient to explain such behavior. Our empirical findings show that deterrence mechanisms (e.g., formal sanctions and whistleblowing expectations) significantly but modestly reduce violation intentions, whereas social-learning mechanisms, especially significant others’ influence and perceived benefits, have stronger effects. These findings highlight the need to understand software developer culture and the social influences, which affect developers’ motivations to abide by the NDA.
LLMs have demonstrated strong language-learning and human-like response-generation capabilities, and they are increasingly used to support decision-making in high-risk sectors. However, their internal decision processes remain difficult to interpret, and their responses may lack transparency. The literature has explored numerous approaches to address transparency challenges in LLMs, including Neurosymbolic AI (NeSy AI). NeSy AI approaches were primarily developed for conventional neural networks and may not transfer directly to the distinctive characteristics of LLMs. Consequently, there is a limited systematic understanding of how symbolic AI can be effectively integrated into LLMs. This paper aims to address this gap by first reviewing established NeSy AI methods and then proposing a novel taxonomy of symbolic integration in LLMs, along with a roadmap to merge symbolic techniques with LLMs. The taxonomy organises the literature across four dimensions: (1) the stage of LLM development at which symbolic information is integrated; (2) the coupling mechanism; (3) the architectural paradigm; and (4) the algorithm-level or application-level perspective. The review identifies commonly used benchmarks, recent advances and important research gaps, and uses these findings to outline directions for future research. By highlighting the latest developments and notable gaps in the literature, it offers practical insights for implementing frameworks for symbolic integration into LLMs to enhance transparency.
Information Systems (IS) discontinuance is an increasingly relevant phenomenon as individuals and organizations continuously reassess their engagement with technology. Research on IS discontinuance has examined this phenomenon across social media, consumer applications, and organizational systems, but the findings remain dispersed across contexts and levels of analysis. This review analyses 97 studies qualitatively using the Gioia method and synthesizes evidence using Leavitt’s socio-technical systems (STS) framework. The analysis identifies drivers and inhibitors across four dimensions: People, Technology, Task, and Structure. It further shows how interactions among these dimensions shape intermittent discontinuance, replacement, and permanent discontinuance. By integrating insights from diverse IS contexts, this review presents discontinuance as a multidimensional process shaped by socio-technical interdependencies rather than a single decision. The findings highlight key theoretical directions for IS discontinuance research and provide practical implications for system designers, organizations, and policymakers managing digital transitions.
Generative artificial intelligence (GenAI) is transforming the travel industry, yet how tourists’ textual inputs influence interactions with GenAI remains underexplored. This study investigates how tourists’ textual inputs influence AI responses and interaction quality. Using a two-study design, we conducted a quasi-experiment to examine subjective interaction quality and a simulation study to examine objective interaction quality. The results show that information density and syntactic complexity increase informational support, whereas subjectivity decreases it. Information density and subjectivity enhance emotional support, whereas syntactic complexity reduces it. The effects on interaction quality differ across dimensions. Information density and syntactic complexity lower subjective but improve objective quality, whereas subjectivity improves subjective quality but has no significant effect on objective quality. Topic consistency further negatively moderates the effects of syntactic complexity on subjective quality and information density on objective quality. These findings advance human-AI interaction literature and offer practical guidance for improving GenAI support for tourists.
Generative Artificial Intelligence (GenAI) is beginning to reconfigure how entrepreneurial work is organized. Professionals with a strong individual entrepreneurial orientation are often among the earliest and most enthusiastic adopters, yet this same zeal can set in motion a chain of effects that may ultimately undermine essential skills. Drawing on cognitive offloading theory and skill decay literature, we conceptualize GenAI use as collaboration depth and GenAI dependence as collaboration asymmetry. We theorize a sequential mechanism in which individual entrepreneurial orientation prompts deeper collaboration with GenAI, sustained collaboration tends to cultivate GenAI dependence, and dependence in turn is associated with perceived erosion of professional skills. We further argue that a strong long-term orientation can amplify the conversion of collaboration depth into dependence, since future-focused individuals may reframe routine use as a strategic investment and assign less immediate priority to corrective practice. Survey data from 143 knowledge workers across five countries are broadly consistent with this account and point to what we term a hustle-to-handicap pattern: when entrepreneurial ambition meets GenAI under a long-term outlook, perceived skill erosion appears to emerge through sustained behavior rather than through disposition alone. The study contributes to the emerging information systems conversation on human-AI collaboration by illuminating how entrepreneurial and temporal orientations shape the conditions under which GenAI may shift from augmenting capability toward perceived capability erosion.
Multi-touch attribution (MTA) aims to assess the effects of various touchpoints on conversions and refine advertising strategies. Recent developments in deep learning have shown substantial advantages in modeling complex user behaviors (i.e., touchpoint sequences) and capturing temporal dependencies in MTA tasks. However, it is unclear about the performance and improvement of deep learning-based MTA models reported in academia. In this study, we present a comparative study of nine MTA models on a synthetic and two public datasets in terms of conversion rate (CVR) prediction and advertising effectiveness. We observe that causal attribution models significantly enhance advertising attribution by addressing the confounding bias induced by user preferences; models with better attribution results perform better in CVR prediction and lead to more efficient advertising decisions (e.g., budget allocation).
This study investigates digital mental health as a high-stakes problem of conversational information access and interaction. Using Design Science Research Methodology, it develops Cognitive Theatre, a risk-aware conversational agent informed by cognitive behavioural therapy and implemented through a controller-mediated role-decomposed architecture. The system separates support into specialised roles for risk assessment, support selection, structured intervention, and user-facing delivery, and coordinates them through risk-aware routing. The artifact was evaluated through safety-routing, latency, AI-based comparative evaluation, and human comparative evaluation. Results show that, under controlled model-generated conditions, the architecture escalated all 50 extreme-risk inputs and received higher response-level ratings than two baselines in AI-based assessment and a two-scenario human evaluation. The study contributes design knowledge for building structured, inspectable, and context-sensitive conversational systems in sensitive domains, and extends research on human-centred conversational information access.
The exponential growth of synthetic media has catalyzed deepfake video detection research, addressing critical concerns regarding content integrity, misinformation mitigation, and privacy preservation. The increasing realism of deepfakes has rendered traditional forensic approaches inadequate for reliably detecting manipulated media. While extensive Generative Adversarial Network (GAN)-based models have emerged, current research remains constrained by significant limitations: (i) GAN-based deepfake models perform poorly under low illumination, occlusion, and information loss; (ii) limited studies address the preservation of 3D geometric details of faces; (iii) current models show weak generalization to unseen faces; and (iv) most deep learning models lack interpretability, limiting transparency and trust. To address these challenges, the present study proposes a new deepfake detection model called Z-numbers-based Granulated Deepfake (Z-GrDeepfake), which leverages facial mask artifacts, PatchGAN discriminator, Fast Fourier Transform (FFT), granulation, and Z-numbers. Unlike state-of-the-art models, Z-GrDeepfake integrates a PatchGAN discriminator with an FFT spectral branch to capture both spatial and frequency features, thereby enabling more robust modeling of face identity information. Furthermore, granulation is performed using blur, illumination, and pose features to enhance attribute retention. Multiple Instance Learning (MIL) aggregation is then employed to detect deepfake images by integrating both identity and attribute information. Finally, Z-numbers are computed based on the deepfake detection score to ensure the reliability in identifying manipulated images. Aforesaid characteristics of Z-GrDeepfake improve its speed and accuracy for deepfake detection. The efficacy of Z-GrDeepfake has been tested on various real and deepfake videos. The superiority of Z-GrDeepfake is also proved through extensive experiments.
Cybercrime rates have increased rapidly during the last decades, resulting in cybercrimes becoming common crimes and leading to the importance of companies’ attention to cybercrime experience. Therefore, this study examines cybercrime experience and provides insights into the companies’ concern, level of information about cybercrime risks and training requirements in SMEs. An empirical analysis is conducted using a dataset collected in Europe in 2021 among more than 12,000 representative respondents of European SMEs. The results show that a higher awareness of cyber risk correlates with fewer cybercrime incidents. However, increased cybersecurity training results in increased reporting of cybercrime incidents. This suggests that increased training improves detection capabilities or that companies’ responses to cybercrime incidents are typically reactive. Increased digitalization is associated with a rise in cybercrime incidents. Therefore, recommended cybersecurity strategies for SMEs include ongoing cybersecurity training programs and the cultivation of a proactive cybersecurity culture to effectively mitigate risks.
Health information systems (HISs) are enhancing healthcare efficiency, but their adoption is slow due to various complex factors. Analysing concerns about HIS adoption from a first-person perspective yields a more nuanced, layered understanding of the underlying reasons. The research utilises the stages of concern (SoC) framework from the Concern-Based Adoption Model (CBAM) to explore individual perceptions and concerns about HIS adoption. The study employs a qualitative abductive research approach, using interviews with 25 healthcare practitioners to analyse the complexities of HIS adoption. The analysis reveals a spectrum of concerns aligned with stages of awareness, information, personal, operational–strategic–stakeholder management, consequences, collaboration, and refocusing. The major impediments to HIS implementation include individuals’ doubts about their personal capability, potential disruptions to existing workflows, and the absence of post-implementation evaluations. The findings highlight the need to consider diverse perspectives to ensure successful technology integration in healthcare.
To examine whether early-stage social signals help explain divergent engagement trajectories beyond video content, this study investigates the lasting impact of the first comment in travel short videos. Utilizing a dataset of 499 matched pairs of cross-posted travel videos and over 400,000 comments from Douyin and Kuaishou, we employ a cross-platform differential modeling approach to reduce content-related confounding and isolate comment-dynamics effects. The findings validate a dual-pathway pattern: (1) a direct sentiment pathway, consistent with emotional contagion, in which the first comment positivity shapes subsequent comment sentiment but attenuates over time; (2) an indirect volume pathway, in which a positive first comment increases early engagement, which in turn predicts higher cumulative comment volume through a persistent herding effect. The study provides a novel lens for understanding commenting dynamics in high-velocity digital ecosystems. For tourism marketers, it suggests differentiated early-stage engagement strategies for short-term sentiment management and longer-term interaction accumulation.
Patient-generated reviews reduce information asymmetry in online health communities (OHCs), as OHCs credence goods markets where service quality is difficult to evaluate. Although prior research has examined static physician attributes, the dynamic communication processes associated with review generation remain underexplored. Furthermore, treating patient feedback as a unidimensional construct overlooks the distinct attributional pathways underlying satisfaction and dissatisfaction. Integrating cue utilization theory with attribution theory, we examine how physicians’ communication cues are asymmetrically associated with positive and negative reviews, and how professional status moderates these relationships. Analyzing 69,954 consultation records using a human-in-the-loop deep learning framework, we identify significant diagnostic asymmetry in cue associations. Emotional support cues show dual diagnosticity, being associated with a 6.9
Fake news increasingly spans multiple languages and modalities, especially on short-form video platforms, making detection more complex. This paper introduces the first study addressing multilingual fake news detection for content that incorporates multiple languages within a single news item. This study utilizes the Design Science Research (DSR) paradigm and presents an automated large language model (LLM)-based fact-checking approach for multilingual fake news detection that requires no extensive task-specific training data. Our method includes: (1) LLM-based cross-modal consistency filtering to identify discrepancies across modalities; and (2) LLM-based fact verification implemented through a multi-agent system. We evaluated our approach using an existing multilingual text fake news dataset and a novel multilingual short-form video fake news dataset (MSVFD). Experimental results show our approach outperforms leading supervised and LLM-based prompting methods. This research addresses a critical gap in multilingual fake news literature and contributes to DSR on LLM-based artifacts, highlighting their potential for timely misinformation detection across diverse formats and languages.
The evaluation of Conversational Recommender Systems necessitates robust protocols to measure utility and user satisfaction. While human-in-the-loop testing remains the gold standard, scalability and reproducibility constraints have driven the field toward User Simulators. However, current simulation paradigms predominantly utilize rigid templates or closed-source Large Language Models that exhibit idealized behaviors. These approaches fail to capture user ambiguity, resulting in benchmarks that overestimate system proficiency by assuming crystallized user intent. To address this limitation, we introduce a family of open-weight user simulation models capable of generalizing across diverse e-commerce domains. Leveraging Teacher-Student distillation, we operationalize three distinct behavioral stereotypes: Direct, Vague-Proactive, and Vague-Reactive. Our evaluation of state-of-the-art Agentic Generative Conversational Recommender Systems reveals a critical Robustness Gap: while agents perform proficiently with decisive users, performance collapses when facing passivity and ambiguity. These findings underscore the necessity of our scalable framework for rigorously stress-testing the next generation of conversational agents against realistic, non-cooperative user behaviors.
As open AI model repositories expand in scale and influence, scalable approaches to evaluate and govern widely adopted models are becoming increasingly essential. This study presents a predictive framework for model popularity, leveraging large-scale data from Hugging Face. Our analysis shows that model adoption is shaped by identifiable features, like task type, language coverage, creator identity, and geographic origin. Despite notable inconsistencies and gaps in model documentation, we achieved over 80
Group holiday planning offers opportunities for shared understanding and stronger social bonds. Yet, it is often difficult to coordinate a group holiday due to conflicting preferences, fragmented information, and the effort required to reach a consensus. Chatbots can help expedite this process by generating options, tailoring recommendations, and accommodating diverse preferences. While prior research has largely examined users’ perceptions of dyadic (single-user) chatbots for holiday planning, the role of chatbots in polyadic, group-based co-creation remains underexplored. To address this gap, we adopt an expectancy–value theory (EVT) perspective to conceptualize collaborative holiday planning with chatbots. We model travelers’ reasons for using a chatbot, such as interaction quality, customization, information quality, and perspective-taking, as well as reasons against using it, including perceived low-quality advice and privacy concerns, as formative predictors of attitudes and, ultimately, willingness to co-create. In a live multi-user chatroom integrating an AI-powered chatbot, 112 participants were organized into 56 pairs, each collaboratively planning a holiday and subsequently reporting their perceptions. Findings indicate that interaction quality, customization, information quality, and especially perspective-taking strengthen reasons for using chatbots, which in turn enhance attitudes and increase willingness to co-create. Conversely, privacy concerns and low-quality advice strengthen reasons against chatbot use, although they do not significantly diminish attitudes. Further, personal values further amplify reasons for using the chatbot. This study extends EVT into the domain of human–AI group collaboration, offering design implications for developing effective polyadic travel-planning chatbots.
