People become more suspicious of a humanoid robot that makes errors, especially when the robot is an expressive conversation partner.
In a study published in Science Robotics, we looked at how 50 people interacted and made joint decisions with Pepper, a commercial humanoid robot designed to express and recognise emotions. Sometimes Pepper offered sound advice. At other times, it made conversational errors, such as interrupting participants or pushing illogical ideas.
For some participants, Pepper was animated, making eye contact, gesturing, and nodding. For others, it remained entirely motionless. We measured four parameters: brain activity, oxytocin levels, self-reported trust, and how much the robot influenced decision-making.
Interestingly, when people interacted with an expressive robot that broke social norms, their oxytocin levels rose. Commonly dubbed the "love hormone" for its role in bonding, oxytocin is usually expected to drop when a partner lets you down. Here, however, higher oxytocin during an animated robot’s errors led to lower trust and fewer accepted suggestions. The hormone was tracking suspicion, not affection.
While errors damaged trust regardless of the robot's expressiveness, an animated presence fundamentally changed how the human brain processed the blunder.
To track real-time brain activity, we used functional near-infrared spectroscopy, a portable forehead sensor that measures brain oxygen levels during natural movement and speech. We focused on two regions: the dorsolateral prefrontal cortex, which flags broken expectations, and the medial prefrontal cortex, which helps infer others' intentions.
When the animated robot erred, participants were caught off guard. Activity spiked across both regions, and they began working together more closely. This synchronised brain activity predicted the rise in oxytocin, which in turn predicted a drop in trust. Conversely, this coordinated brain response was entirely absent when participants interacted with expressionless robots.
As robots enter homes, hospitals, and workplaces, user adoption hinges heavily on trust. Designers often assume that lifelike, expressive traits build a reservoir of trust that protects a robot's reputation when it fails.
Our findings challenge this assumption. Expressive cues appear to transform a mistake from a minor technical malfunction into a social violation. A motionless robot’s error looks mechanical, but the same error from an animated robot triggers the brain mechanics used to judge human behaviour.
This study was limited to young men using a single robot model, and our sensor only tracked the front of the brain. Future research must test whether this oxytocin-linked vigilance occurs in women, diverse age groups, across different cultures, and with other robot designs.
We also plan to investigate whether robots can rebuild lost trust after a blunder by acknowledging the error, apologising, or signalling good intent, just as humans do after an awkward social exchange.