As people increasingly turn to AI for personal deliberation beyond task-oriented assistance, concerns about sycophancy in these value-laden contexts have grown. Unlike human flattery, which is intentional and self-interested, AI sycophancy emerges as a byproduct of RLHF’s reward structure for user-preference alignment. Yet the observable behavior is similar: both produce responses that preserve what users want to hear. Focusing on this phenomenon through Goffman’s face-work framework, we operationalize AI sycophancy as excessive face-saving, either active (preserving positive face through agreement) or passive (preserving negative face by withholding challenge). In a mixed-methods study (N=31), participants engaged with AI across three moral dilemmas under these conditions and a non-sycophantic neutral baseline. Sycophantic responses increased decision confidence but reduced open-minded thinking; participants felt supported yet found the conversations unproductive. Neutral responses, though initially uncomfortable, promoted cognitive flexibility and meaningful deliberation. These findings reveal a confidence-competence trade-off in AI-mediated moral reasoning and suggest that effective AI for personal deliberation requires calibrated friction, not unconditional agreement.

Generative information retrieval (GenIR) enables users to obtain synthesized information through iterative interaction with LLMs, fundamentally reshaping how AI-generated content is produced and consumed. Within this shift, users may encounter AI-generated informational content through two primary pathways: actively creating it themselves or consuming content generated by others. We examine whether authorship biases evaluation---whether users judge AI output from their own interactions more favorably than equivalent output from others. In a mixed-methods experiment (N=28, 2×2 within-subjects), participants interacted with an AI system to retrieve and craft information, then evaluated both their own result and equivalent output generated through the same process but framed as someone else's. Results reveal a selective authorship bias: participants rated self-obtained information significantly higher in quality, but showed no corresponding difference in trust. This pattern suggests that hallucination-aware skepticism constrained trust judgments, but could not prevent quality-driven selection behavior, even in the presence of information conflicts. Given that iterative interactions are inherent to GenIR, diverse interventions seem needed to support users' critical evaluation.

Keyeun Lee, Seo Hyeong Kim, Seolhee Lee, Jinsu Eun, Yena Ko, Hayeon Jeon, Esther Hehsun Kim, Seonghye Cho, Soeun Yang, Eun-mee Kim, Hajin Lim. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

With the advancement of AI, even people without professional experience can create artworks using AI-based image generation systems like DALL-E 2. However, little is known about how users interact with these new AI algorithms, much less how AI-infused systems can be designed. We explore the user experience of these new technologies and their potential to foster creativity. A user study was carried out where 13 participants executed tasks of creating artworks using DALL-E 2 alongside in-depth interviews related to their experience. The results showed that users had ambivalent opinions regarding the algorithm’s performance. When users were informed of the system’s capabilities, they subsequently utilized more specific prompts to generate the intended output. Users also optimized their prompts (the queries they entered to create artworks) based on how algorithms worked to achieve their desired outcome. The users wanted a two-way interaction where AI explained the outcome and accepted feedback rather than simply accepting unilateral instructions. We discuss the implications for designing interfaces that maximize creativity while providing comfort for the users.

Large language models (LLMs) enhance productivity and creativity, but many users struggle to formulate appropriate prompts, discouraging consistent usage. We introduce PromptPilot that assists users by recommending context-appropriate prompts based on task types and the user input. We evaluated PromptPilot through an online experiment using a 3 × 3 mixed factorial design. The study involved 273 participants and examined three initiative conditions (AI-initiative, mixed-initiative, user-initiative) as a between-subjects variable, across three distinct task types (browsing, daily ideation, brainstorming) as a within-subjects variable. We found that the AI-initiative and mixed-initiative systems yielded superior performance results compared to the user-initiative system. Notably, participants in the mixed initiative generated prompts using fewer words compared to those in the AI and user-initiative. The proportion of AI-generated prompts in the AI-initiative was 2.3 times that of the mixed-initiative. We discuss implications for user interaction where AI can support users’ prompting process.

Students with disabilities (SWDs) often struggle with note-taking during lectures. Therefore, many higher education institutions have implemented peer note-taking programs (PNTPs), where peer note-takers (PNTs) assist SWDs in taking lecture notes. To better understand the experiences of SWDs and PNTs, we conducted semi-structured interviews with eight SWDs and eight PNTs. We found that the interaction between SWDs and PNTs was predominantly unidirectional, highlighting specific needs and challenges. In response, we developed EmoBridge, a collaborative note-taking platform that facilitates real-time collaboration and communication between PNT-SWD pairs using emojis. We evaluated EmoBridge through an in-the-wild study with seven PNT-SWD pairs. The results showed improved class participation for SWDs and a reduced sense of sole responsibility for PNTs. Based on these insights, we discuss design implications for collaborative note-taking systems aimed at enhancing PNTPs and fostering more effective and inclusive educational experiences for SWDs.

Online chat functions as a discussion channel for diverse social issues. However, deliberative discussion and consensus-reaching can be difficult in online chats in part because of the lack of structure. To explore the feasibility of a conversational agent that enables deliberative discussion, we designed and developed DebateBot, a chatbot that structures discussion and encourages reticent participants to contribute. We conducted a 2 (discussion structure: unstructured vs. structured) × 2 (discussant facilitation: unfacilitated vs. facilitated) between-subjects experiment (N = 64, 12 groups). Our findings are as follows: (1) Structured discussion positively affects discussion quality by generating diverse opinions within a group and resulting in a high level of perceived deliberative quality. (2) Facilitation drives a high level of opinion alignment between group consensus and independent individual opinions, resulting in authentic consensus reaching. Facilitation also drives more even contribution and a higher level of task cohesion and communication fairness. Our results suggest that a chatbot agent could partially substitute for a human moderator in deliberative discussions.

Artificial intelligence (AI) algorithms are making remarkable achievements even in creative fields such as aesthetics. However, whether those outside the machine learning (ML) community can sufficiently interpret or agree with their results, especially in such highly subjective domains, is being questioned. In this paper, we try to understand how different user communities reason about AI algorithm results in subjective domains. We designed AI Mirror, a research probe that tells users the algorithmically predicted aesthetic scores of photographs. We conducted a user study of the system with 18 participants from three different groups: AI/ML experts, domain experts (photographers), and general public members. They performed tasks consisting of taking photos and reasoning about AI Mirror's prediction algorithm with think-aloud sessions, surveys, and interviews. The results showed the following: (1) Users understood the AI using their own group-specific expertise; (2) Users employed various strategies to close the gap between their judgments and AI predictions overtime; (3) The difference between users' thoughts and AI pre-dictions was negatively related with users' perceptions of the AI's interpretability and reasonability. We also discuss design considerations for AI-infused systems in subjective domains.

Although group chat discussions are prevalent in daily life, they have a number of limitations. When discussing in a group chat, reaching a consensus often takes time, members contribute unevenly to the discussion, and messages are unorganized. Hence, we aimed to explore the feasibility of a facilitator chatbot agent to improve group chat discussions. We conducted a needfinding survey to identify key features for a facilitator chatbot. We then implemented GroupfeedBot, a chatbot agent that could facilitate group discussions by managing the discussion time, encouraging members to participate evenly, and organizing members' opinions. To evaluate GroupfeedBot, we performed preliminary user studies that varied for diverse tasks and different group sizes. We found that the group with GroupfeedBot appeared to exhibit more diversity in opinions even though there were no differences in output quality and message quantity. On the other hand, GroupfeedBot promoted members' even participation and effective communication for the medium-sized group.