Believing you’re working with a machine lowers your confidence — even if you’re right

A recent study published in Neuroscience of Consciousness sheds light on a surprisingly powerful factor that shapes how we make — and revise — decisions: our own sense of confidence. The research found that people’s subjective confidence in their perceptual judgments more strongly predicted whether they would change their mind than how accurate their decision was or how difficult the task seemed. Moreover, confidence dropped when people believed they were interacting with a machine instead of another person, even when both “partners” performed identically.

The researchers wanted to better understand the role of confidence in decision-making. While many theories propose that confidence is a key internal signal that helps guide learning, behavior, and interactions with others — including machines — there have been relatively few direct studies testing exactly how it influences decisions. One challenge is that confidence and accuracy tend to go hand in hand, making it difficult to tease apart which one truly drives behavior. The study set out to address that gap by designing a task that could separate the influence of confidence from task accuracy and difficulty.

“Most studies on confidence — the evaluation of our own decisions — focus on understanding how confidence feelings are generated, specifically the underlying signals that influence why we feel more or less confident. However, less attention has been given to the actual function of confidence in cognition,” said study author Rémi Sanchez of ONERA and Aix-Marseille University.

“Why do we experience confidence as a distinct feeling? Does it serve a purpose in decision-making? In this study, we aimed to explore whether confidence plays a role in guiding interactions with a partner during a perceptual task. We made the participants believe that their partner could be either another human who had previously performed the task, or a machine trained to perform as well as a human. Would the nature of the partner, a human or a machine, impact participants’ decision-making behavior?”

“Additionally, we wanted to examine whether participants’ confidence can be tracked through eye pupil dynamics. While a few studies have investigated this, the findings remain inconclusive, so we sought to bring further clarity to this question.”

For their study, the researchers recruited 14 adults to complete a perceptual decision-making task. Participants viewed two displays of moving dots and had to decide which one moved in a direction closer to vertical. After making this initial choice, they rated how confident they were in their answer. Then, they were shown the response of a supposed partner — either another human or a computer algorithm — and were told they could either keep their original answer or switch to match their partner’s.

Importantly, both “partners” were actually computer-generated and were programmed to perform equally well across all trials. This allowed the researchers to manipulate only the perceived identity of the partner — machine or human — while keeping everything else constant.

The task was designed to encourage participants to vary their confidence. The researchers adjusted the difficulty of the motion discrimination task for each participant and also manipulated how consistent the motion direction was within each dot display — a feature known to influence confidence without affecting actual performance. Eye-tracking equipment was also used throughout the experiment to measure changes in pupil size, which can reflect shifts in arousal and attention.

Across hundreds of trials, participants were more likely to change their answer when they felt unsure about their initial decision. In fact, confidence predicted whether someone would switch their answer almost three times more strongly than whether their initial answer was actually right or wrong. People changed their minds in about half of the trials where they felt low confidence, compared to only 16% of trials where their answer was incorrect — an indication that how sure they felt mattered more than whether they were right.

“Confidence predicted participants’ switch decisions (change of mind) more strongly than difficulty or accuracy,” Sanchez told PsyPost.

Interestingly, this tendency to rely on confidence was influenced by social context. Participants reported feeling less confident in their decisions when they believed they were interacting with a machine, even though their actual performance didn’t change. They also switched their answers more often in these “machine partner” trials than in the human partner trials. These effects suggest that context — in this case, who you think you’re working with — can influence internal evaluations like confidence, which in turn shape how we behave.

Importantly, the partner’s identity had no impact on how accurate participants were in their initial responses. This suggests the drop in confidence when interacting with a machine wasn’t driven by worse performance or more difficult stimuli, but likely reflected a belief that machines might be more reliable — despite being told that both human and machine partners performed the same.

“We were quite surprised to find that participants had lower confidence in their own decisions when they believed their task partner was a machine — even though their actual performance remained the same,” Sanchez explained. “It is not clear why, but we speculate that, in our study, participants may have assumed the machine was inherently better at the task, despite being explicitly told that it performed at the same level as a human. Lowering their own confidence might have been a strategic adjustment, making them more inclined to rely on the machine’s decisions in an attempt to maximize overall task performance.”

Beyond behavior, the researchers also found a link between confidence and physiological signals. By analyzing changes in pupil size before participants made their responses, the researchers discovered that pupil dilation predicted how confident a person would later report feeling. In general, greater dilation — specifically, a faster rate of change — was associated with higher confidence. This effect occurred even before the decision was made, suggesting that subtle changes in arousal or attentional state might be tied to how confident we feel.

Although the ability to predict confidence from pupil dilation was modest, it supports the idea that physiological cues might provide a way to monitor internal states like confidence in real-time — a useful tool in settings where people can’t explicitly report how they feel.

“Confidence judgments could be slightly predicted by pre-response eye pupil dynamics, indicating that arousal changes are linked to confidence computations,” Sanchez said.

The study also examined eye blinks, which can influence or accompany pupil changes. Blinks were more common in trials where participants were incorrect, felt less confident, or changed their answer — potentially reflecting increased cognitive effort or shifts in attention. This suggests that eye behavior in general could offer insight into internal decision-making processes.

The study, “What the eyes, confidence, and partner’s identity can tell about change of mind,” was authored by Rémi Sanchez, Anne-Catherine Tomei, Pascal Mamassian, Manuel Vidal, and Andrea Desantis.