9월 4주차

Housing discrimination has been recognized as an important societal issue for decades. While this issue can manifest in multiple ways, one of the most observed avenues is price discrimination, where houses in white-dominant neighborhoods are worth more than houses in minority-dominant neighborhoods that are otherwise similar. Prior studies have empirically documented such pricing discrimination and attributed it to human biases. In addition, recent studies have shown that issues of this kind are unlikely to be addressed by traditional AI models, even those specifically designed to address discrimination. In this paper, we first compare AI-generated versus human-generated housing prices using a sample of 284,749 U.S. properties. We then study the impact of generative AI in the context of price discrimination in the housing market and find that it can help alleviate this issue. Our mechanism exploration provides empirical evidence regarding underlying mechanisms that drive such a counter-intuitive result. Practical and policy implications are also discussed.

Privacy-related news can embroil firms in public controversy, eroding user trust and triggering negative word-of-mouth. We examine whether firms can mitigate downstream damage before a crisis unfolds through the new approach we propose—active privacy transparency. Unlike conventional practices that leave information about firms’ data practices buried in lengthy privacy policies, active privacy transparency proactively presents precise information about a specific data practice when users make the corresponding permission decision, enabling a genuinely informed choice. Across behavioral experiments, we find that this preventive approach mitigates trust decline and negative word-of-mouth when users later encounter negative third-party news about the firm’s privacy practices. Importantly, these downstream benefits do not necessarily come at the immediate cost of lower permission granting. Rather, the immediate effect depends on users’ privacy knowledge, and more knowledgeable users may even become more willing to grant permissions. For managers, these findings position active privacy transparency as a preventive privacy-management approach rather than a postincident remedy. For policymakers, these findings call for moving beyond broad transparency mandates toward clearer guidance on what information should be provided and when, where, and how it should be communicated, while supporting initiatives that strengthen users’ privacy knowledge.

Platforms are increasingly deploying engagement rewards, granting users redeemable benefits for everyday activity such as logging in, commenting, and sharing content. Platform managers nevertheless hesitate, because a reward that increases viewer engagement may have unintended consequences for creator earnings, and managers rarely know in advance where it will bite. We study one such setting, video-game live-streaming, where viewers can convert engagement-earned coins into free virtual gifts that enter the same voluntary paid gifting channel that streamers rely on for income. A free sticker gives the same on-screen recognition as the cheapest 10-cent paid sticker, so viewers no longer need to pay to be seen appreciating the streamer. Intuition says free gifts could cannibalize paid ones. We find the opposite: access to free gifts increased paid gift spending by roughly $2.42 per viewer-streamer relationship per week. The reason is that free rewards dilute the signaling value of the cheapest paid gift, and committed viewers upgrade to costlier gifts to stand out. Free rewards complement monetization when gifting is publicly visible in real time and paid tiers remain clearly differentiated from free options. For policymakers, the question is what this does to creator income. Engagement rewards subsidize one side of a two-sided market, and the effect on the other side is rarely documented. Voluntary gifts from viewers are creators’ primary income, so a token program that hands viewers free versions of these gifts could erode it. Our evidence comes from an exogenous policy change on a major platform, and creators' paid gift income rose rather than fell. Free gifts also converted previously non-paying viewers into paying supporters, and the gains held across streamer popularity tiers, so the policy did not concentrate benefits among already-popular creators. Token programs designed this way need not trade off engagement growth against creator compensation.

Organizations invest heavily in digital platforms, yet sustaining user engagement after initial adoption remains a persistent challenge. In feature-rich services, such as mobile banking, customers’ usage evolves as they gain experience, develop routines, and encounter changing needs. However, existing research largely treats postadoption information systems (IS) usage as a stable outcome, offering limited insight into how usage evolves over time. Drawing on longitudinal mobile banking usage data, surveys, and qualitative evidence, this study develops and validates a mechanism-based process model of postadoption feature usage dynamics. We show that feature-level usage evolves through three dominant mechanisms: evaluation-driven entry, habit-driven stabilization, and need-driven disruption and re-evaluation. To capture these dynamics, we frame IS usage as comprising usage intensity (frequency and depth) and usage length, and we identify five distinct longitudinal usage trajectories (e.g., expanders and maximizers). We demonstrate that these trajectories are associated with important organizational outcomes, including new feature adoption, monetization willingness, loyalty, and security compliance. Thus, organizations should move beyond simply increasing usage frequency and instead, recognize that users follow different developmental paths requiring different engagement strategies. This work provides managers with a framework for identifying users’ usage trajectories and designing targeted interventions that better support long-term customer engagement and value creation.

Design-based causal methods such as difference in differences, instrumental variables, randomized controlled trials, regression discontinuity, and synthetic control have become the foundation of evidence-based decision making in information systems (IS). Their value, however, depends on the credibility of the published evidence. Analyzing 7,516 hypothesis tests from 558 articles published in seven leading IS journals between 2000 and 2023, we identify systematic clustering of statistical results immediately above the conventional 5% significance threshold, particularly for difference in differences, instrumental variables, and randomized controlled trials. These patterns suggest that publication incentives and researcher degrees of freedom may distort the evidence base used by managers and policymakers. We translate these findings into practical recommendations for researchers, reviewers, and editors. Specifically, we advocate greater transparency in research design and reporting, stronger requirements for data and code availability, broader acceptance of well-executed null findings, routine sensitivity analyses, and method-specific reporting standards for causal studies. We also introduce an IS-calibrated local false discovery rate that helps readers assess the credibility of reported causal estimates. By strengthening evidence quality rather than discouraging causal research, these recommendations aim to improve the reliability, reproducibility, and policy relevance of empirical findings that increasingly guide organizational investment, digital transformation, platform governance, and public policy decisions.

While conventional software is programmed for specific tasks, AI learns and refines its capabilities through a two-stage process: pre-training and fine-tuning. A central feature of this process is the AI adaptive learning embedded in the fine-tuning stage, whereby user interactions refine the AI’s capabilities initially developed from publicly available data during pre-training. This feature reshapes the firm’s freemium decision. We develop an analytical model to examine how AI adaptive learning affects firms’ optimal freemium strategies and find that while a free version can expand the user base and generate learning data, it also intensifies cannibalization of demand and increases the service-cost burden. We demonstrate that highly effective adaptive learning can paradoxically harm profitability by excessively enhancing the free version’s appeal and sharply cannibalizing premium demand. We find further that firms may optimally avoid offering a free version even when service costs are minimal, relying instead on reductions to the price of premium service to expand the paid user base. Narrowing the base capability gap between versions can also increase profitability by strengthening adaptive learning despite the increased cannibalization. Finally, results show that introducing a free version does not necessarily increase consumer surplus, because improved AI capabilities may allow firms to raise premium prices. These results are robust across several model extensions and inform both firms’ monetization strategies and policy discussions of consumer welfare in generative AI markets.

We theorize that designing conversational agents (CAs) to appear more humanlike will make minor errors appear more normal because to err is human. When errors appear more normal, users are less likely to strive to identify their cause (a process called attribution) and thus are less likely to respond negatively. We conducted two experiments to test our theoretical model, and the results generally support our theorizing: greater perceived humanness preserves the perception of situational normality when an error occurs, thereby reducing error attribution and mitigating the negative effects of errors on service satisfaction. Our research contributes to the theory by identifying a theoretical mechanism that underlies users’ responses to errors (a reduction in situational normality triggers error attribution). It also has important implications for practice by showing that designing CAs to be more humanlike is important for CAs more likely to make errors (e.g. CAs controlled by large language models) and less important for other CAs (e.g. those controlled by robust rule-based scripts).

Prevailing information privacy theories have advanced understanding of how users evaluate privacy risks. Yet, they leave important aspects of user behavior insufficiently explained, particularly why some consumers proactively protect their personal information, whereas others remain passive despite expressing similar privacy concerns. Much of this research relies on privacy concerns as the primary indicator of privacy attitudes. Privacy concerns are crucial, but they largely capture users’ risk-focused evaluations of data collection, use, and disclosure. They therefore explain how users react to perceived privacy threats better than how users develop sustained engagement in protecting their privacy. We propose and test the Theory of Privacy Interests to explain this consumer-side engagement. By privacy interests, we mean users’ experiential engagement with protecting their personal information—the extent to which privacy protection is meaningful to them, they feel competent to pursue it, and they believe their actions can influence privacy outcomes. This construct refers to users’ privacy-protective interests, not to the economic or strategic interests of firms, platforms, or other producers that supply privacy-adjacent digital environments. Drawing on the Heideggerian onto-epistemological framework, we conceptualize privacy interests as experience-based, skillful engagement with privacy protection that develops through users’ repeated interactions with digital environments, privacy risks, and privacy-protective practices. We empirically examine privacy interests and privacy concerns as distinct but complementary constructs. Across three scale-development data collections and four survey studies with 3,922 participants, we find that privacy interests are more effective in explaining proactive privacy actions. In contrast, privacy concerns are more effective in explaining reactive privacy actions. This research operationalizes the Heideggerian onto-epistemological framework to shift attention from users’ risk evaluation to users’ sustained engagement in actualizing privacy protection. Our research offers a complementary explanation for variation in consumer privacy behavior and advances research on privacy attitudes, privacy actions, and the privacy paradox.

This paper examines the use of large language models (LLMs) for human-LLM co-creation, wherein humans use LLMs to accomplish text-based tasks requiring knowledge and understanding of a topic. Task factuality, the correspondence of the human-LLM co-creation task output to reality and verifiable facts, is an important outcome of such tasks, yet is difficult to achieve. We investigate how the human’s reflection enhances task factuality in such tasks. Theorizing two aspects of reflection, namely, the human’s cognitive state and the interaction mode with the LLM, we develop hypotheses explaining how: (1) two types of interaction modes (adversarial and conversational) differentially enhance task factuality; and (2) three types of cognitive states (shallow, dialogic and critical) mediate the differential effect of interaction mode on task factuality. We test our hypotheses through a randomized experiment on a task in which participants wrote a short essay on a specific topic by working with an LLM. Integrating data from experimental manipulations (interaction mode), survey measures (cognitive state) and objective assessment (task factuality and cognitive state) drawn from 280 LLM users, the paper makes theoretical contributions by: explaining how reflection can enhance epistemic integration between humans and LLMs by increasing task factuality in human-LLM co-creation tasks, theoretically unpacking the concept of reflection in the context of human-LLM co-creation, and providing insights for LLM design that can lead to higher factuality of such tasks. Practical implications for LLM users are to engage in reflection when working with LLMs to generate more factual outputs, for organizations to develop employee capacity for reflection, and for LLM companies to design features that foster reflection for users.

On online platforms, new ideas often emerge by recombining existing ones within idea networks. Unlike traditional knowledge networks, idea networks represent curated, meaning-based associations among ideas, offering a distinct lens on recombination. Drawing upon a hypergraph perspective, we investigate how new idea success depends on their structural and content attributes, and how collaborative participation shapes these attributes. Using data from an ideation platform, we find that both structural embeddedness and bridging benefit new idea success. Content diversity has no direct effect, but it amplifies the benefits of bridging while constraining those of embeddedness. Both crowd contributions and ideator expertise strengthen ideas’ structural positions, whereas they shape content diversity in opposite ways: crowd contributions increase diversity, while ideator expertise reduces it through selective integration. These findings advance research on networks, recombination, and online collaboration by showing how structure, content, and collaborative participation jointly shape new idea success.

The quality of online review content is a central factor in its perceived helpfulness to consumers. Yet platforms’ efforts to enhance the quality of online reviews through extrinsic incentives have shown mixed results. In this paper, we address this issue by showing that a subtle change in the design of the review generation process can improve review quality. Via four randomized controlled experiments (n = 1,101), we specifically examine whether the task order in which platforms solicit numerical ratings and textual reviews can shape reviewers’ reasoning processes and, in turn, the deliberativeness of textual reviews and the characteristics of subsequent numerical ratings. Results show that changing the task order in the online review generation process, from the current “rating first” to a “review first” task order, enhances reviewers’ deliberation in their textual reviews, reduces their ambivalence toward the product/service, and improves their decisiveness and confidence in numerical ratings. Moreover, we show that these effects are contingent on the device (PC vs. mobile) used to generate textual reviews, as a situational boundary condition. In response to calls for research on the design of review systems, these findings offer actionable insights for online review platforms by presenting novel ways to elicit more deliberative textual reviews and more decisive, confident numerical ratings. They also clarify the important distinction between PC reviews and mobile reviews.

Generative AI (GenAI) has advanced rapidly and made significant impacts. However, its effect on developers remains a topic of industry debate. Companies want to know whether GenAI can enhance developers’ coding performance, as an unclear understanding may put companies at a disadvantage. While the literature has begun addressing this issue, a formal understanding of GenAI’s impact remains incomplete. Moreover, existing findings are often short-term, fragmented, or lack explanatory mechanisms. To fill these gaps, we designed a multimethod research program comprising a longitudinal field study and a randomized controlled experiment. In Study 1, we collaborated with a global information technology organization and applied a difference-in-differences approach to over 27 weeks of proprietary data. In Study 2, we designed a randomized experiment involving 253 software developers. From these studies, we find that GenAI usage affects both developers’ coding quantity and quality. These effects, however, depend critically on how the tool is used. While reduced cognitive effort can be associated with diminished quality, interestingly, GenAI usage enables developers to produce higher-quality code with less cognitive effort. In this current study, we explain the paradoxical findings through cognitive load theory, showing that GenAI reduces extraneous load while preserving germane processing during ideation and debugging. Using a multimethod research design that integrates longitudinal field data with a randomized controlled experiment, we link observed performance effects to underlying cognitive mechanisms and usage strategies. We also offer guidance on effective usage styles and propose boundary conditions for realizing GenAI’s benefits in practice.

This research examines how startup complementors’ product security influences their funding from external investors in platform ecosystems. The longitudinal analysis of the Hadoop ecosystem shows that the release of product security features is positively associated with startup funding. The effect is stronger when public media pays more attention to the platform ecosystem’s security issues and for complementors with a central position in the platform ecosystem’s technological proximity network. We further find that the effect depends on the strength and interpretability of the product security signals. Specifically, the effect is driven by product-embedded security efforts rather than by partnership-based initiatives, and is more pronounced among platform-native complementors. These findings advance our understanding of how product security functions as a strategic signal in ecosystems characterized by interdependent risks and networked competition. We offer practical implications for complementors and platforms seeking to translate security investments into competitive advantage and strengthen security incentives.

Artificial intelligence (AI) developers have leveraged open source software (OSS) to accelerate AI’s progress. However, this has introduced security issues, including newly developed machine learning open-source software (MLOSS) repositories inheriting vulnerabilities from each other, typically lacking any explicit signal. In this study, we adopted the computational design science paradigm to design a novel MLOSS Link Prediction framework to map the spread of vulnerabilities across AI. We propose a Self-Supervised AI-Feature Aware Graph Attention Autoencoder (SSAIF-GATE) to learn from a sparsely labeled network, a novel AI-Feature Aware attention mechanism that captures shared AI terms, and a multilevel pretext task to leverage multiple components of a network’s structure. SSAIF-GATE outperforms prevailing graph embedding methods with an area-under-the-curve of 94.8 percent and an average precision of 96.1 percent. SSAIF-GATE helps address extensive vulnerability spread among MLOSS and contributes design principles that can inform future information technology artifact design for broader domains including business intelligence and healthcare.

Although opportunities are central to firm innovation and performance, prior research lacks a scalable, theory-grounded approach to measuring them. Existing measures are either context-specific or detached from explicit relational mechanisms, limiting their generalizability and interpretability. Leveraging computational methods including text analytics, unsupervised machine learning, and network analysis of large-scale digitized data, we propose a computational design framework guided by structural hole theory that enables fine-grained strategic opportunity measures: hole-opening, hole-entering, and non-hole positions. We validate this framework through systematic analysis of initial public offering (IPO) outcomes using U.S. public firm panel data. The results show that hole-opening positions are associated with higher post-IPO valuations, but with a lower likelihood of mergers and acquisitions (M&A) exits, whereas hole-entering and non-hole positions are linked to lower IPO valuations but higher probabilities of M&A outcomes. These patterns reveal distinct opportunity roles based on firms’ relative structural positions. This computational framework contributes to IS research by offering a replicable, theory-driven foundation for opportunity measurement.

Modeling and leveraging consumers’ dynamic search behaviors presents significant business opportunities. Although deep learning methods excel at processing vast consumer data for predictive tasks, their opaque nature limits interpretability and fails to explicitly model consumer decision making. In contrast, economic theory suggests that consumers follow a sequential search strategy, evaluating alternatives until they find the best match for their preferences. To bridge this gap, we propose the sequential search transformer (SST), a deep structural econometric model that integrates deep learning with sequential search theory to model search and purchase decisions. SST unifies these two approaches into an end-to-end trainable model, improving both predictive accuracy and policy evaluation capabilities. Unlike conventional deep learning models, SST explicitly models consumer decision making, and unlike existing sequential search models, it enables consumer behavior modeling across sessions and sequentially resolves utility uncertainty for searched items. We provide a theoretical analysis of the identification strategy for the SST model and show that all parameters can be identified under the proposed framework. Then, we apply SST to a data set with detailed clickstream data collected from a U.S. e-commerce website. Empirical evaluations show that SST outperforms state-of-the-art deep learning and structural models in predicting consumer searches and purchases. Moreover, policy experiments demonstrate SST’s effectiveness in optimizing product recommendations and new product promotion strategies, ultimately enhancing consumers experience and driving revenue growth. This paper was accepted by Hemant Bhargava, information systems. Supplemental Material: The online appendices and data files are available at https://doi.org/10.1287/mnsc.2024.04540 .