Abstract: This study examines how the adoption of standalone large language model services, specifically ChatGPT, alters users’ commercial search behavior on traditional search engines and their engagement with e-commerce platforms. Using Nielsen Korea panel data from November 2022 to April 2023, we track user activity across Google, Bing, Naver, Daum, and major online retailers. Grounded in a purchase-funnel framework, we conceptualize ChatGPT as a cognitive amplifier that enables users to translate vague, exploratory needs into precise, goal-oriented queries. Applying a propensity-score-matched difference-in-differences design, we find that adopters of ChatGPT increase their commercial search volume by 31.7 percent and diversify their query mix by 29.9 percent, with a notable shift toward greater semantic clarity and specificity. Two-stage least-squares analyses confirm that these behavioral changes generate substantial increases in page views and visit durations across both search engine shopping portals and retail sites. Our findings advance the literature on digital platforms and AI-mediated decision-making by showing how generative AI tools reshape search behavior in ways that support, rather than replace, traditional search engines. These results offer clear implications for multiple stakeholders. Search engine developers should prioritize the integration of conversational AI for real-time query refinement. Marketers should adapt content strategies to align with more intent-rich and natural language queries. Policymakers and industry leaders must also consider how the convergence of LLMs and traditional search systems may influence market structure, user experience, and competition in the evolving digital economy. 

Abstract: The success of long-term service programs, such as education, weight loss, debt management, and rehabilitation, relies heavily on customer adherence to service provider instructions. However, customers are often noncompliant to instructions due to the unpleasant and challenging nature of achieving long term goals. This paper examines whether substituting human service providers with AI can potentially improve customer compliance and enhance service outcomes. Using data from a natural experiment setting in South Korea, our empirical findings revealed that AI does not contribute to customer noncompliance in the same way human providers do. While human providers are more effective with compliant customers, AI excels with noncompliant ones. Based on these findings, we propose a strategic division of labor, where AI handles noncompliant customers, allowing human providers to focus on compliant ones for better service outcomes. In an era of generative AI, our findings also challenge the common belief that humans always outperform AI in services requiring deep interpersonal interactions.

Abstract: Generative AI, such as ChatGPT, is anticipated to reshape information search behavior, with prior research highlighting its potential to reduce biased information acquisition and polarization. Based on a large-scale, year-long longitudinal and cross-platform observation of individuals’ ChatGPT adoption and search behaviors in the highly polarized political context, this study empirically investigates whether and how ChatGPT influences information search behavior and examines its implications for (political) learning, (de)polarization, and knowledge construction at both individual and collective levels. The findings indicate that, at the individual level, ChatGPT adoption increases overall search activities and even facilitates exploration of previously unused queries, fostering seemingly ideologically balanced search behavior. However, the study also finds that generative AI consistently directs users toward popular, widely used search terms at the population level. Popular terms also happen to be ideologically less extreme, suggesting a cautious interpretation of AI’s debiasing or depolarization effects. The variance-convergence paradox—individual-level search diversification coupled with collective-level convergence toward popular terms in search queries—challenges optimistic interpretations of AI’s potentials based solely on individual-level outcomes from one-shot experiments and suggests future research consider broader, longer-term, collective-level implications of AI. 

Abstract: We examine the impact of free sample promotion within the changing retail landscape, characterized by a substantial pricing gap between online retailers and manufacturer-owned direct channels. We demonstrate that free samples increase both page views and visit duration, both of which are key indicators for purchase. Yet, the data also reveals a 16.1% decrease in monthly spending. Based on subsequent analyses, we suggest a potential scenario where the observed spending reduction might signify consumers transitioning their purchase channel from direct channels to retail channels. Our findings contribute to the research on free sample promotion by moving beyond merely assessing whether customers purchased the promoted product or not, and instead examining the specific purchase channel through which the customer made their purchase. Our findings are increasingly relevant in a context where manufacturers advocate the use of direct channels in order to reduce pricing gaps and avoid excessive reliance on retail channels. 

Abstract: Mobile devices have evolved as a major channel for online video consumption. With the growth of online video platforms, various ad formats have been experimented to increase advertising effectiveness. One such attempt is pre-roll skippable ads that allow users to skip an ad after watching it for a few seconds. Despite pre-roll ads providing better user experience, their high skipping rate is alarming to advertisers and brands. Recent studies revealed what factors influence pre-roll ad acceptance, yet little has been known regarding the role of device type. This study investigates how mobile users respond to pre-roll skippable ads differently than PC users, focusing on individual and contextual factors of video consumption. Using individual-level clickstream data obtained from a major online video platform in South Korea, this study found that mobile users tend to skip pre-roll ads more often than PC users, but the degree differs by individual and contextual factors. 

I examine how to improve safety guardrails in generative AI—an essential component of product design— by balancing two competing goals: helpfulness and harmlessness. Political bias in AI poses legal, regulatory, and reputational risks. In mitigate these, industry practice typically enforces neutrality by presenting “both sides” of controversial issues, assuming users will perceive such responses as fair. I argue, however, that excessive guardrails can backfire. Drawing on motivated reasoning theory, I hypothesize a gap between technical neutrality and perceived neutrality: users may interpret neutral responses as biased when they contradict their beliefs, leading to backlash and altered behavior. In collaboration with Kakao, one of South Korea’s largest tech firms, I will test these hypotheses through an experiment. This study contributes to a long-standing literature on perceived algorithmic bias and trust. Prior research has found similar effects in domains such as news media (Hostile Media Effect) and social platforms. To my knowledge, it is the first empirical study examining how safety design in generative AI influences real-world user choice. 

Agentic AIs are capable of autonomously completing multi-step tasks such as trip planning or restaurant booking. These tasks are often initiated through underspecified user instructions and require the AI to infer, clarify, or supplement missing information. This creates a fundamental trade-off: while prompting users with clarifying questions can reduce service failures by filling in gaps, it also imposes cognitive effort and risks frustrating users or causing task abandonment. Striking the right balance is difficult, as users vary widely in their tolerance for back-and-forth interaction. To address this challenge, we propose a simple but powerful intervention: a user-facing toggle that allows individuals to decide whether the AI should “ask more questions” or simply “take it from here.” We are currently designing an experiment to evaluate how this form of adaptive delegation influences user satisfaction, blame attribution, and willingness to delegate future tasks—across both successful and failed service outcomes. Grounded in principal-agent theory, we hypothesize that giving users control over interaction depth improves satisfaction by resolving a core tension between cognitive effort and perceived agency. This choice should reduce dissatisfaction and blame when outcomes are poor, and enhance perceived value when outcomes are successful—particularly for users with a higher need for cognitive engagement. 

Research Question:  How does giving users an ‘Ask-More ↔ Proceed’ toggle change their satisfaction, blame attribution, and future delegation intent when (a) the agent fails vs. (b) succeeds only after many clarifications?