An Expressive Humanoid That Errs Breeds Distrust, Study Finds

An expressive humanoid that fumbles a conversation draws more distrust than the same robot standing still. That is the finding of a study published on 31 August 2026 in Science Robotics, in which 50 participants talked with a Pepper robot while their brain activity was recorded. Gestures and eye contact make us read a mistake as a social violation rather than a mechanical fault.

Key takeaways

The study covered 50 participants, all of them young men
Brain activity was measured with forehead-mounted fNIRS, which allowed free movement
The recorded regions were the dorsolateral prefrontal cortex and the medial prefrontal cortex
An expressive robot breaking conversational norms raised the listener’s oxytocin levels
Higher oxytocin went together with lower trust and less influence over the participant’s decisions

How the experiment worked

Participants held conversations and made joint decisions with a Pepper robot from SoftBank Robotics. For some of them the robot gestured, nodded and held eye contact. For the rest it stayed motionless. Either way it sometimes gave sound advice and sometimes cut in mid-sentence or suggested something illogical. Brain activity was recorded throughout with fNIRS, a sensor mounted on the forehead.

The study was carried out by Hasan Ayaz and Yigit Topoglu of Drexel University, Frank Krueger of George Mason University and Ewart J. de Visser of the United States Air Force Academy.

The study design in four points:

50 participantsyoung men only — the main limitation of the sample

fNIRSa forehead sensor measuring cortical oxygenation, usable during free movement

Peppera single robot model, run in two modes: expressive and motionless

oxytocina biochemical marker sampled alongside the brain measurement

Expressiveness raises the stakes

The dorsolateral prefrontal cortexdorsolateral prefrontal cortex: A brain region that responds to uncertainty and to situations where someone breaks an expected norm. responds to uncertainty and norm violations. The medial prefrontal cortex handles inferring other people’s intent. When the expressive robot erred, both regions activated in a coordinated way. With the motionless robot that effect was absent.

Put differently: the more a machine behaves like a conversational partner, the more its slip-ups get judged in human terms.

Oxytocin does not mean warmth

The most surprising result concerns oxytocinoxytocin: A hormone associated with bonding and trust. In this study its rise tracked vigilance rather than affection., usually associated with bonding and trust. Here its rise accompanied the opposite: the higher the level, the lower the trust in the robot and the smaller its sway over the participant’s choices. The authors link this to social vigilance rather than affection.

Why it matters

The humanoid industry treats expressiveness as a free path to acceptance: more facial motion, more gesture, warmer reception. This study suggests the reverse — expression takes on a debt that the robot only repays through flawless behaviour. For designers of service robots that is a concrete instruction: either reliability matches the promise written on the face, or do not promise that face at all.

What the study does not say

The sample was 50 young men only, so the results cannot be generalised to the wider population. It also tested a single robot, Pepper, rather than humanoids with different facial mechanics.

What's next?

The paper is available in Science Robotics under DOI 10.1126/scirobotics.aec1762
The authors attribute the effect to social vigilance — the mechanism needs confirmation on a broader sample
It remains open whether the effect holds for robots with different facial mechanics