Researchers have found that people quickly develop trust in expressive, engaging humanoid robots, but that trust can collapse just as rapidly when the robots make mistakes.

The study led by researchers at Drexel University is the first to simultaneously measure brain activity, hormone levels, self-reported attitudes, and behavior during human-robot interactions.

By tracking both psychological and biological responses, the researchers gained new insights into how trust is built and lost while collaborating with humanoid robots.

The findings could help designers create more reliable social robots for healthcare, education, customer service, and home environments.

Expressive robots falter

Researchers have shown that making humanoid robots more expressive can strengthen human engagement, but it also makes trust far more fragile when the robots make mistakes. The new study explored how people respond to a humanoid robot named Pepper during extended face-to-face interactions.

Unlike previous studies that relied mainly on questionnaires, the team simultaneously monitored participants’ brain activity using wearable functional near-infrared spectroscopy (fNIRS), measured oxytocin levels from saliva samples, tracked behavioral decisions, and recorded subjective trust ratings. This multimodal approach enabled researchers to observe how neurological, physiological, and behavioral responses evolved in real time during conversations with a robot.Fifty adult participants interacted with one of two versions of Pepper. One robot displayed human-like social behaviors, including eye contact, head nods, hand gestures, and verbal acknowledgments such as “uh-huh.” The second robot delivered the same scripted dialogue but remained motionless and provided no nonverbal cues.

Each participant completed three interaction sessions. The first two were error-free to establish trust, while the third deliberately introduced socially disruptive behaviors. Instead of experiencing mechanical failures, participants encountered conversational errors such as irrelevant responses, interruptions, illogical reasoning, and requests to repeat themselves. The expressive robot initially generated stronger engagement than the motionless version. Participants paid more attention to its responses and were more willing to consider its recommendations during collaborative decision-making tasks. However, this advantage disappeared once the robot began making mistakes.

Brain recordings showed that errors made by the expressive robot activated the dorsolateral and medial prefrontal cortex—regions involved in social reasoning and interpreting other people’s intentions. Researchers concluded that participants processed these failures as violations of social expectations rather than simple technical glitches. The motionless robot produced a different response. Participants appeared to interpret its mistakes as isolated system errors instead of flaws in a social partner, resulting in weaker neural reactions and a less dramatic decline in trust.

Reliable humanoids win

The study also produced an unexpected physiological finding. Oxytocin, commonly associated with bonding and social connection, increased after the robots made mistakes even though trust declined. The researchers suggest that, in human-robot interactions, oxytocin may function as a vigilance signal that heightens attention to unreliable social behavior rather than indicating stronger emotional attachment.

Behavioral data reinforced the neural findings. Once the expressive robot started making errors, its ability to influence participants’ decisions dropped by more than half. Although both robot versions experienced sharp declines in trust, expressive behaviors amplified the consequences of failures instead of protecting against them.

The findings highlight an important engineering challenge for social robotics. Human-like gestures, eye contact, and conversational cues can improve engagement, but they also raise users’ expectations for reliability. As humanoid robots are increasingly deployed in healthcare, education, customer service, and home environments, designers may need to carefully balance social expressiveness with robust, dependable performance, since even small conversational mistakes can significantly undermine user trust.

“Robot design shouldn’t look at likeability in isolation. Likeability matters, but engineering a social robot must take into account neurobiology and psychology to maximize performance,” said Ewart J de Visser, of the Warfighter Effectiveness Research Center at the US Air Force Academy and a corresponding author on the study, in a statement.