When a robot breaks your trust, your brain knows it, George Mason University collaborative study finds
The first human-robot interaction (HRI) study ever to track human trust in a robot across brain, hormone, mind, and behavior, at the same time, finds that the more engaging robot had the most to lose.
As artificial intelligence becomes increasingly woven into everyday life, researchers are racing to answer a deceptively simple question: What happens when people start trusting machines the way they trust other people?
A new study published in Science Robotics offers one of the most comprehensive answers yet, revealing that when a socially engaging robot makes mistakes, people don't simply see a technical failure. Trust is affected, and they experience something that looks remarkably like a social betrayal.
“This novel brain-based research partnership brings together neural activity, hormones, psychology, and behavior to examine how people develop and lose trust in socially interactive robots,” explained Frank Krueger, a neuroscientist and trust expert at George Mason University. Krueger contributed equally to the research with first author Yigit Topoglu, who received his doctorate from Drexel University this year.
Topoglu, who now works at the Warfighter Effectiveness Research Center at the U.S. Air Force Academy, emphasized the study’s live interaction approach. “By observing participants during face-to-face interactions with the robot, we could examine not only what people reported about their trust, but also how their brains, hormones, and behavior changed when the robot began violating their expectations,” Topoglu said.
“Our team’s findings extend beyond neuroscience, psychology, and human-robot interaction,” Krueger continued. “They also have important implications for healthcare, education, defense and national security, customer service and hospitality, and consumer technology—particularly for those designing socially interactive robots, embodied AI systems, and AI assistants and deciding how human-like these technologies should be.”
The research brought together investigators from Drexel University, George Mason University, the U.S. Air Force Academy, the University of Southern California's Institute for Creative Technologies, and other collaborators. Their goal was ambitious: track trust in a robot simultaneously through brain activity, hormone responses, psychological assessments, and actual behavior during live, face-to-face interaction.
“Trust is not one thing you can capture with a single measurement, so we measured several at once,” said Hasan Ayaz, a professor in Drexel’s School of Biomedical Engineering, Science, and Health Systems and the study’s senior author. “We recorded brain activity, hormone levels, what people told us on surveys, and what they actually did—all in the same session, while a person sat across from the robot and talked with it. Each modality tells a different part of the story, and only together do they show how trust builds and how it breaks down.”
Ewart J. de Visser, of the Warfighter Effectiveness Research Center at the U.S. Air Force Academy and a corresponding author on the study, emphasized the practical importance of the findings. “In real-world human–robot teams, trust must be calibrated rather than simply maximized. A system that appears highly social may encourage stronger engagement, but it may also produce a greater loss of trust when its behavior violates people’s expectations.”
Fifty adult men participated in the study, spending roughly two and a half hours interacting with Pepper, a humanoid robot. At first, the conversations and collaborative tasks unfolded smoothly. Participants chatted with the robot about topics ranging from travel to music, and worked together on decision-making exercises.
Then things changed.
During a later interaction, the robot began to make mistakes. It interrupted participants, offered illogical reasoning, drifted off topic, and violated basic social norms. Importantly, the errors were not mechanical breakdowns. They were social missteps.
The researchers also created two versions of Pepper. One version was expressive, making eye contact, nodding, gesturing, and providing verbal cues that mimicked human conversation. The other remained still and expressionless while delivering the exact same words.
That distinction turned out to matter substantially.
Participants were more engaged by the expressive robot. Yet when that robot began behaving poorly, trust declined faster and more dramatically than it did with the motionless version.
“A charming robot that slips up pays a steeper price than a plain one,” said Krueger. “Expressiveness is not free. It buys you engagement, and it buys you fragility at the same time, and that's a trade-off designers should be making deliberately rather than by accident.”
What makes the study particularly novel is that researchers could observe what was happening beneath the surface. Using wearable functional near-infrared spectroscopy (fNIRS), they monitored activity in participants' prefrontal cortex during the interactions. They also measured oxytocin levels from saliva samples, collected trust surveys, and analyzed whether participants actually followed the robot's advice.
The findings revealed a striking chain of events.
When the expressive robot violated social expectations, activity increased in regions of the brain associated with social reasoning. At the same time, oxytocin levels rose. Yet rather than signaling increased bonding, higher oxytocin was linked to declining trust and reduced willingness to follow the robot's recommendations.
For decades, oxytocin has often been described as the "love hormone" because of its role in human bonding and relationships. In this case, however, it appeared to function more like a warning signal.
The researchers suggest that participants interpreted the expressive robot's failures as interpersonal violations rather than technical glitches. Their brains responded as though a social partner had let them down.
The implications extend far beyond robotics laboratories.
As AI-powered assistants, healthcare technologies, educational tools, customer-service platforms, and autonomous systems become more conversational and human-like, developers face a critical design question: How social should AI be?
The study suggests there is a trade-off. Human-like behaviors can strengthen engagement and rapport, but they also raise expectations. When an expressive AI system fails, people may react not as users confronting a malfunctioning machine, but as individuals responding to a broken social relationship.
And in a future increasingly shaped by human-AI collaboration, understanding that difference could prove essential.
Funding for this research was provided by the U.S. Department of Defense Air Force Office of Scientific Research through the Minerva Research Initiative under grant #FA9550-18-1-0455.
