Why a friendlier robot loses your trust faster when It messes up

Fifty men sat down with Pepper, the humanoid robot known for its expressive gestures and eye contact, and talked through decisions with it. Sometimes Pepper gave solid advice. Sometimes it interrupted, pushed illogical suggestions, or otherwise broke the unwritten rules of conversation. Half the participants got an animated version of the robot, nodding and gesturing like a person. The other half got a version that just sat there, motionless, delivering the same words with none of the body language.

Researchers led by Hasan Ayaz and colleagues, publishing in Science Robotics, tracked brain activity with a forehead-worn sensor, measured oxytocin levels, and logged how much people actually trusted and followed the robot's advice. The twist showed up in the hormone data. Oxytocin usually gets branded the bonding chemical, the thing that rises when you feel close to someone. Here it did the opposite job. When the expressive robot screwed up, oxytocin spiked, and the more it spiked in a given person, the less they trusted the robot afterward and the less likely they were to take its advice.

The brain scans explain why. Two regions, the dorsolateral prefrontal cortex and the medial prefrontal cortex, lit up and started syncing more tightly when the animated robot made a mistake. One tracks broken expectations, the other handles reading intentions, the same circuitry people use to figure out what a confusing friend or coworker meant. That coordinated activity predicted the oxytocin rise, which predicted the trust collapse. None of this happened with the motionless robot. Its errors just registered as glitches, not social slights.

That distinction matters more than it sounds. A robot that gestures and holds eye contact isn't just decorated software, it gets processed by the same mental machinery reserved for people, and mistakes get reinterpreted as betrayals rather than bugs. Robot makers have generally assumed that lifelike design buys goodwill, a kind of buffer against errors. This study suggests the buffer doesn't exist, and might run in reverse.

The researchers are upfront about the limits: all-male sample, one robot, one culture, and a brain sensor that only reached the front of the skull. Their next move is testing whether apologies or acknowledgment from the robot can repair the trust hit, the way a sincere sorry works between two humans. Given how fast companies are shipping expressive robots into homes and care facilities, that follow-up can't come soon enough.