Responsiveness in large language model (LLM) applications is widely assumed to be critical, yet the impact of latency on user behavior and perception of output quality has not been systematically explored. We report a controlled experiment varying time-to-first-token latency (2, 9, 20 seconds) across two taxonomy-driven knowledge task types (Creation and Advice). Log analyses reveal that user interaction behaviors were robust to latency, yet varied by task type: Creation tasks elicited more frequent prompting than Advice tasks. In contrast, participants who experienced 2-second latencies rated the LLM’s outputs less thoughtful and useful than those who experienced 9- or 20-second latencies. Participants attributed delays to AI deliberation, though long waits occasionally shifted this interpretation toward frustration or concerns about reliability. Overall, this work demonstrates that latency is not simply a cost to reduce but a tunable design variable with ethical implications. We offer design strategies for enhancing human-LLM interaction.
Systems' latency — the time between user input and system response — slows down the human-computer interaction loop. Several studies revealed negative objective and subjective effects of high latency, typically treating latency as a constant delay. Because latency varies significantly in practice, recent work also assessed the effects of large and sudden latency changes. In practice, however, latency variations are small but frequent. As the effects of such variations are unclear, we investigate how small latency variations (+/- 50 ms) affect users' performance and perceived task load for 2D target selection tasks with static and moving targets. For static targets, we found that latency variation causes significantly higher completion times and less efficient trajectories, however with small effect sizes. In contrast, we found no significant effects on any performance measure for moving targets. Our findings indicate that the effect of latency variation is generally very small and quickly disappears for non-trivial tasks.
Users feel frustrated when they do not know when to speak with LLM-based agents. Technical delays disrupt the natural rhythm of conversation (turn-taking), yet there is little understanding of how these specific delays impact the back-and-forth flow of interaction. To address this, we analyzed human-agent conversations in social VR to measure timing differences. We used conversation analysis techniques to track specific timing metrics, such as how long it takes to respond (response latencies) and how agents handle interruptions (repair attempts). We found that agents are significantly slower to respond with a median of 4.1 seconds compared to a human's 1.2 seconds. We identified a "conversational timing drift", noting that agents struggle with start-up latency, i.e., taking too long to start speaking, and wind-down latency, i.e., failing to stop speaking quickly when a user interrupts them. This is the first study to empirically quantify human-agent conversational latencies within VR. We offer design suggestions to help future agents manage conversational timing better, ultimately improving natural conversation and user experience.
Collaboration is a key use case for conversational AI agents. Yet we know little about how agents' collaborative effort affects users' reciprocal effort and how this relates to perceptions of agent conversational capability. Through an online director-matcher task (n = 267) whereby participants interacted with agents that varied in their collaborative effort, we found that users rated agents that were less communicatively collaborative as less competent partners. Yet, contrary to the division of labour principle in communication, users only increased their own collaborative effort as speakers when communicating with more collaborative agents, whilst also benefitting more as listeners when interacting with such agents. We discuss the implications of these findings bringing together partner modelling and division of labour principles in driving human-agent collaborative communication in both speaker and listener effort, and consider the strategic application of agent collaborative effort in the design of conversational AI.
Recent Large Language Model (LLM) based AI can exhibit recognizable and measurable personality traits during conversations to improve user experience. However, as human understandings of their personality traits can be affected by their interaction partners' traits, a potential risk is that AI traits may shape and bias users' self-concept of their own traits. To explore the possibility, we conducted a randomized behavioral experiment. Our results indicate that after conversations about personal topics with an LLM-based AI chatbot using GPT-4o default personality traits, users' self-concepts aligned with the AI's measured personality traits. The longer the conversation, the greater the alignment. This alignment led to increased homogeneity in self-concepts among users. We also observed that the degree of self-concept alignment was positively associated with users' conversation enjoyment. Our findings uncover how AI personality traits can shape users' self-concepts through human-AI conversation, highlighting both risks and opportunities. We provide important design implications for developing more responsible and ethical AI systems.
System-imposed wait times can significantly disrupt digital workflows, affecting user experience and task performance. Prior HCI research has examined how temporal feedback, such as feedback mode (Elapsed-Time vs. Remaining-Time) shapes wait-time perception. However, few studies have investigated how such feedback influences users’ downstream task performance, as well as overall affective and cognitive experience. To study these effects, we conducted an online experiment where 425 participants performing a visual reasoning task experienced a 10-, 30-, or 60-second wait with a Remaining-Time, Elapsed-Time, or No Time Display. Findings show that temporal feedback mode shapes how waiting is perceived: Remaining-Time feedback increased frustration relative to Elapsed-Time feedback, while No Time Display made waits feel longer and heightened ambiguity. Notably, these experiential differences did not translate into differences in post-wait task performance. Integrating psychophysical and cognitive science perspectives, we discuss implications for implementing temporal feedback in latency-prone digital systems.
Large Language Models (LLMs) are increasingly used to generate and edit scientific abstracts, yet their integration into academic writing raises questions about trust, quality, and disclosure. Despite growing adoption, little is known about how readers perceive LLM-generated summaries and how these perceptions influence evaluations of scientific work. This paper presents a mixed-methods survey experiment investigating whether readers with ML expertise can distinguish between human- and LLM-generated abstracts, how actual and perceived LLM involvement affects judgments of quality and trustworthiness, and what orientations readers adopt toward AI-assisted writing. Our findings show that participants struggle to reliably identify LLM-generated content, yet their beliefs about LLM involvement significantly shape their evaluations. Notably, abstracts edited by LLMs are rated more favorably than those written solely by humans or LLMs. We also identify three distinct reader orientations toward LLM-assisted writing, offering insights into evolving norms and informing policy around disclosure and acceptable use in scientific communication.