As robots become increasingly common in hospitals, schools, customer service and people’s homes, one question is becoming more important: what makes people trust a robot?
A new study published in Science Robotics suggests that the answer is more complicated than simply making robots appear friendlier or more human.
Researchers found that people engaged more deeply with a humanoid robot that made eye contact, nodded, gestured and gave brief verbal signals to show that it was listening. However, this greater social engagement also appeared to raise expectations.
When the robot began interrupting people, making irrelevant comments and offering illogical arguments, trust and behavioral influence declined. These effects occurred with both an expressive and a largely motionless version of the robot, but the expressive robot produced a distinctive pattern connecting brain activity, salivary oxytocin, reported trust and people’s willingness to follow its suggestions.
The findings suggest that human-like behavior may make interactions with robots feel more natural, while also making their mistakes feel more socially significant.
Why trust matters
Trust plays an important role whenever people rely on technology.
A navigation app needs to provide reliable directions. An artificial intelligence assistant needs to give accurate information. A robot supporting nurses, teaching children or assisting older adults must also earn people’s confidence if it is to be accepted.
Engineers have therefore tried to make social robots more approachable by giving them human-like qualities. Some robots can maintain eye contact, use gestures, nod while another person is speaking and offer small acknowledgements such as “uh-huh” to show that they are listening.
These behaviors can make a machine appear more attentive and socially aware. But the new research indicates that greater expressiveness does not automatically protect a robot from losing trust when something goes wrong.
Two versions of the same robot
The study involved 50 healthy adult men who participated in laboratory sessions lasting roughly two and a half hours.
During each session, the participants spoke face-to-face and completed collaborative tasks with Pepper, a humanoid robot frequently used in human-robot interaction research.
Although the participants believed they were interacting directly with the robot, Pepper was secretly controlled by a human operator using a method sometimes called a “Wizard of Oz” experiment. The operator followed a prepared script so that the researchers could keep the robot’s words and mistakes consistent across the different sessions.
The participants were divided between two versions of Pepper.
The expressive version maintained eye contact, moved its arms while speaking, nodded as participants talked and occasionally used brief verbal acknowledgements to signal that it was listening.
The stationary version used the same spoken words but showed very little movement or non-verbal behavior.
This allowed the researchers to study whether the robot’s physical and conversational expressiveness changed how people reacted to its behavior.
Building trust and then introducing errors
Participants first completed two interactions in which the robot behaved reliably.
The conversations covered familiar topics such as food, travel, music and imaginary superpowers. Participants also completed collaborative activities, including selecting useful items for survival on a deserted island and discussing works of art.
During these early interactions, the robot’s comments were relevant and generally logical. This gave participants an opportunity to develop expectations about how it would behave.
During a later interaction, the robot began making deliberate social and conversational errors.
It interrupted participants, asked them to repeat themselves, made comments unrelated to the discussion and defended its opinions using illogical arguments.
These were not dramatic mechanical failures such as freezing, losing power or shutting down. Instead, they resembled the kinds of mistakes a person might associate with a distracted, rude or inconsiderate conversation partner.
Greater engagement raised the stakes
The expressive robot initially encouraged deeper engagement. Its eye contact, gestures and listening signals made the interaction feel more socially responsive than the interaction with the stationary robot.
However, trust declined after both versions began making mistakes.
The most important difference was not simply that one robot was trusted and the other was not. Instead, expressiveness changed the relationship between what participants experienced, what they reported and how they behaved afterward.
In the expressive condition, lower trust was more closely connected to changes in prefrontal brain activity, salivary oxytocin concentrations and participants’ willingness to change their decisions after hearing the robot’s advice.
According to a summary published by Drexel University, the robot’s measurable influence over participants’ decisions fell by more than half after it began making errors.
The researchers interpreted the expressive robot’s effect as evidence that people may process mistakes differently when a machine behaves like a social partner.
Rather than experiencing the errors only as technical failures, participants may have viewed them as violations of the expectations created by the robot’s human-like behavior.
This remains an interpretation rather than direct proof that people experienced the robot’s behavior in exactly the same way they would experience betrayal or rudeness from another person.
Looking at the brain, hormones, trust and behavior
The study was designed to examine trust at several levels instead of relying only on questionnaires.
The researchers used four main types of measurement.
First, participants wore a lightweight brain-imaging device that used functional near-infrared spectroscopy, or fNIRS, to monitor changes in blood oxygenation in the prefrontal cortex.
This part of the brain is involved in decision-making, attention and social reasoning.
Second, the researchers collected saliva samples to measure oxytocin concentrations.
Third, participants completed questionnaires describing how much they trusted the robot.
Finally, the researchers examined behavior by measuring whether participants changed their choices after receiving the robot’s recommendations.
Combining these measurements allowed the research team to compare what people said about the robot with what they actually did and with the biological changes observed during the experiment.
The authors describe the work as the first study to combine neural, hormonal, psychological and behavioral measures simultaneously during direct social interaction with a humanoid robot.
The complicated role of oxytocin
One of the study’s most notable results involved oxytocin.
Oxytocin is sometimes described as the “bonding hormone” because it has been associated with social relationships, attachment and trust. However, scientists increasingly recognize that its effects depend heavily on context.
In this experiment, salivary oxytocin concentrations increased after the robot began making errors, even though reported trust declined.
In the expressive condition, higher oxytocin was also associated with lower trust and reduced willingness to follow the robot’s advice.
The researchers suggest that, in this context, the increase may have reflected heightened vigilance or social sensitivity rather than stronger emotional bonding.
In other words, oxytocin may have been connected to participants paying closer attention to the expressive robot’s unexpected and socially inappropriate behavior.
The study does not establish that oxytocin directly caused people to become less trusting. It identifies an association between hormone levels, trust and behavior that will need to be examined in further research.
Reliability still matters most
The expressive robot was initially more engaging, but its social behavior did not prevent trust and influence from falling once it repeatedly made mistakes.
The findings support a practical lesson for robot designers: human-like behavior can create stronger expectations.
When a robot uses eye contact, gestures and natural conversational signals, people may expect it to behave like a competent and considerate social partner. If it then interrupts them, ignores the topic or makes irrational arguments, the mismatch may become especially noticeable.
A less expressive machine may create fewer social expectations. Its errors can still reduce trust, but people may be more likely to process them as limitations of the technology rather than as violations of an established social relationship.
This does not mean robot designers should avoid expressiveness. Gestures, eye contact and listening signals can make interactions easier and more comfortable.
It does suggest that social features need to be supported by reliable performance. A robot that appears socially intelligent but behaves inconsistently may create disappointment that a more obviously mechanical system would not.
Why the findings matter
Social robots are being developed for a growing range of roles, including supporting healthcare workers, assisting older adults, helping in classrooms, welcoming customers and providing information in public spaces.
In many of these settings, a robot’s usefulness depends partly on whether people believe its recommendations and feel comfortable interacting with it.
The study suggests that designers should think about trust as something that changes over time rather than as a single first impression.
An expressive robot may make a strong initial impression, but that advantage can become a liability if its behavior does not remain dependable.
Designers may therefore need to consider how robots acknowledge mistakes, explain uncertainty and repair trust after an interaction goes wrong.
A robot that recognizes an error, apologizes clearly or explains the limits of its knowledge may be received differently from one that continues behaving confidently after providing a poor or irrelevant response.
The current study did not test those possible trust-repair strategies, but they represent an important direction for future research.
Important limitations
Like all laboratory research, the study has limitations.
All 50 participants were healthy adult men, so the findings may not apply equally to women, children, older adults or people from different cultural and social backgrounds.
Each experimental condition therefore included a relatively small number of participants.
The research also involved one type of humanoid robot, a controlled laboratory setting and a scripted set of conversations. People may respond differently to autonomous robots operating in hospitals, schools, homes or other real-world environments.
Several types of social error were introduced together, including interruptions, irrelevant statements and illogical arguments. The study therefore cannot show which individual mistake had the greatest effect on trust.
In the main experimental sequence, the robot’s unreliable interaction came after two reliable interactions. This means that the effects of the errors cannot be separated completely from the passage of time or the order of the sessions.
The researchers conducted additional validation analyses involving different ordering and greater separation between the interactions. These produced results that were generally consistent with the main findings, although the reported effects were comparatively small.
The results should therefore be treated as detailed evidence from a controlled experiment rather than as proof that every expressive robot will produce the same response.
Looking ahead
As robots become more capable and enter more areas of everyday life, engineers will face a delicate balancing act.
Natural gestures, eye contact and conversational signals can help people interact with machines more comfortably. At the same time, these features can encourage people to judge robots according to higher social standards.
The new research suggests that expressiveness does more than make a robot appear friendly. It may change how people interpret the robot’s errors and how closely trust is connected to their later decisions.
For developers of social robots, the message is straightforward: expressiveness may help begin the relationship, but consistent and reliable behavior is essential for maintaining it.
Citation
Topoglu, Y., Krueger, F., Joshi, S., Rothstein, N., Franke, A. A., Li, X., Gratch, J., de Visser, E. J., & Ayaz, H. (2026). Multilevel dynamics of the brain, hormones, mind, and behavior in social human-robot interaction. Science Robotics, 11(116), Article eaec1762. https://doi.org/10.1126/scirobotics.aec1762
