Neurophysiological Dynamics of Trust in Human–AI Interaction: A Multi-Level Study of Brain, Hormone, Mind, and Behavior
As artificial intelligence (AI) increasingly participates in human decision environments, understanding how trust emerges and breaks down in human–AI interaction has become a central challenge for information systems research. While prior work has predominantly focused on behavioral measures, the underlying neurophysiological mechanisms of trust remain largely unexplored. In this study, we investigated the multi-level dynamics of trust during face-to-face interaction with an embodied intelligent agent. Participants engaged in decision-making tasks with a humanoid robot while neural activity was recorded using functional near-infrared spectroscopy, salivary oxytocin levels were assessed, and self-reported trust and behavioral influence were measured. We experimentally manipulated system reliability (congruent vs. erroneous decisions) and social expressiveness (animated vs. stationary behavior). Results demonstrated that reliability constitutes the primary foundation of trust: agent errors significantly reduced both reported trust and behavioral influence. Social expressiveness modulated these effects. Animated agents elicited stronger prefrontal activation and enhanced neural–hormonal coupling. Notably, elevated oxytocin levels were associated with reduced trust and diminished behavioral influence when expressive agents committed errors, indicating a context-sensitive vigilance response rather than a simple affiliative bonding mechanism. Together, these findings establish a multi-level neurophysiological framework for understanding trust in human–AI interaction and reveal a critical design trade-off between social expressiveness and trust robustness in intelligent systems.
Dr. Frank Krueger is Professor of Systems Social Neuroscience at George Mason University, where he leads the Social Cognition and Interaction: Functional Neuroimaging (SCI:FI) Lab and serves as Core Faculty at the Center for Advancing Systems Science and Bioengineering Innovation. He is also Honorary Professor of Psychology at University of Mannheim, Germany. Trained in psychology, neuroscience, and physics, his research focuses on the psychoneurobiological foundations of trust in human–human and human–AI interactions. His work integrates behavioral science, social neuroscience, and neuroergonomics to understand how trust emerges, adapts, and can be designed in increasingly autonomous socio-technical systems. Dr. Krueger has authored over 240 publications and edited The Neurobiology of Trust with Cambridge University Press. He is Specialty Chief Editor for Frontiers in Social Neuroergonomics and Founder and President of the TRUST Foundation (Transdisciplinary Research Union for the Study of Trust), advancing global, transdisciplinary collaboration on trust in the age of AI.
Behaviour Change Agents in Healthcare: From Individual Nudges to Systemic Impact
Healthcare faces unparalleled pressure from aging populations, workforce shortages, and fragmented, multimodal data streams. How can we transform this data overload into trustworthy, actionable clinical decisions? This talk explores the transformative potential of behaviour change agents in healthcare such as robotics and digital twins, moving beyond simple physiological models to create workflow-aware decision infrastructures. By synthesizing technical design, AI integration, human factors, and clinical impact, we investigate two distinct case studies: RehabTwin, a behavior change agent for home-based stroke rehabilitation and a multi-scale behavior change agent for acute cardiovascular intervention planning. Through these contrasting settings, from the high-stakes lab to the patient's living room, I will highlight some recent evaluation outcomes and implementation barriers, and I look forward to engaging with the audience to discuss future challenges and the broader impacts on our field.
Prof. Ben Allouch is a scholar in human–robot interaction (HRI), advancing data-driven interaction design. Her research integrates human-centered methodologies and real-world clinical validation to optimize robot-assisted care. She bridges applied and fundamental research to shape responsible and societally embedded human–robot interaction.
The motivation to learn from tutors in the context of human robot interaction
In the context of human-agent interaction where the robot or the human can play the role of a coach, we explore the levers of motivation, and in particular of intrinsic motivation. The computational models of this psychological theory allow automatic curriculum learning, thus automatic scaffolding adapted to each learner. The learning program is thus adaptive to the progress of individuals. We illustrate through examples of a robot coach for physical rehabilitation and interactive learning robots learning hierarchical skills.
Sao Mai Nguyen specializes in cognitive developmental learning, reinforcement learning, human-in -the-loop learning, intrinsic motivation, automatic curriculum learning for robots, assistive robotics, activity modelling : she develops algorithms for robots to learn multi-task controls by designing themselves their curriculum and by actively requesting human demonstrations, focusing on long-horizon activities.
She received her PhD from Inria in 2013, holds an Engineer degree from Ecole Polytechnique, France and a master’s degree in adaptive machine systems from Osaka University, Japan. She is currently a professor at Ensta IP Paris, France and was previously with IMT Atlantique, France. She also acts as an associate editor of the journal IEEE RA-L and IJRR and the chair of the Task force “Action and Perception” of the IEEE Technical Committee on Cognitive and Developmental Systems.
Response Generation for Motivational Interviewing: Dialogue Strategies toward Behavior Change
Motivational Interviewing (MI) is a collaborative style of communication designed to elicit a client's own motivations for behavior change. Generating counselor responses in MI has recently attracted growing attention in dialogue generation research, yet producing responses that genuinely guide clients toward change remains a challenge. In this talk, I will review recent studies on response generation in MI and highlight the importance of dialogue management in eliciting clients' own statements in favor of change ("change talk"). To address this challenge, I will present our schema-based approach to dialogue strategy decision, in which the system dynamically determines the focus of each response in a manner grounded in MI principles. I will also discuss evaluation methods for assessing how well generated responses align with MI principles and skills. Finally, I will reflect on the risks of over-reliance on AI in counseling communication and what they imply for the responsible design of such systems.
Yukiko Nakano is a Professor in the Faculty of Science and Technology at Seikei University. She received her M.S. in Media Arts and Sciences from Massachusetts Institute of Technology and Ph.D. in Information Science and Technology from the University of Tokyo. Motivated by the goal of enabling more natural human–computer interaction, she has worked on modeling conversations through the analysis of human verbal and nonverbal communicative behaviors. Her research interests include social signal processing for estimating the characteristics of multimodal and multiparty interactions, and applying these empirical models to human-agent interactions, such as conversational agents and communication robots. She has served as co-chair and senior program committee member for major international conferences on interaction and intelligent agents, including IUI, ICMI, AAMAS, and IVA.
When agents move us: Empathy beyond explanation and persuasion
Agents increasingly influence how people think, decide, and act, making the design of trustworthy behavior-changing interactions an important challenge. Recent research has focused heavily on explanation, persuasion, personalization, and trust, yet human responses to agents are also shaped by social and relational processes that are not fully captured by these perspectives. In this talk, I focus on empathy as a lens for understanding how people come to accept and respond to agent influence. I first review recent developments in human–agent interaction and behavior-change research, and then introduce examples from our studies on self-disclosure, empathic behavior, trust, explanation, persuasion, and responsibility. These findings illustrate how agent influence can emerge through multiple pathways, including subtle interactional cues and social evaluation. Finally, I discuss how future agent design might move beyond maximizing persuasive effectiveness toward supporting influence that is understandable, acceptable, and compatible with human agency.
Takahiro Tsumura is an Assistant Professor at the Faculty of Information Networking for Innovation and Design (INIAD), Toyo University, Japan. His research focuses on human–agent interaction, with particular interest in empathy, trust, responsibility attribution, and behavior change in interactions with AI agents and robots. He studies how people accept, rely on, and emotionally respond to agents in situations where explanations are limited, ambiguous, or difficult to interpret. His recent work explores the design of agents that can support human judgment and shared responsibility beyond explicit explanation, including the role of empathy and nonverbal or non-semantic cues in agent-mediated decision making.
Data-Driven Steering in Human–AI Collaborative Decision-Making: Potential and Problems
In this talk, I will introduce our previous attempts to steer user decisions based on data-driven user models and discuss their potential and problems. In this framework, the user model takes the task context, including cues provided by the AI, as input and predicts the user’s final decision. By using the model to predict how different AI cues affect the user’s decision, the AI can identify the cue that is most likely to steer the user toward a particular decision. Our previous work suggests the potential that, given a sufficiently accurate model, user decisions could be steered in this way. At the same time, it also raises several concerns: users’ responses to cues intended to steer them may differ substantially depending on their personality, and steering users toward an inaccurate target decision may instead degrade their decision-making. Based on these possibilities, I will discuss how AI-based steering should be designed.