Trial-to-trial decision shifts shape confidence, study finds

For decades, confidence has been treated as a mental afterglow of decision-making: the stronger the evidence seems, the more certain people feel. A new study challenges that seemingly straightforward picture. Researchers report that confidence is not determined only by the strength of sensory evidence or by how far an internal decision signal lies from a fixed boundary. Instead, confidence also reflects the momentary state of the decision criterion itself—a psychological threshold that can shift from one trial to the next. The findings suggest that what appears to be random variation in confidence may contain meaningful information about how the brain is evaluating evidence in real time.

The study, published in Nature Human Behaviour, brings together theoretical analysis, computational modelling and data from 15 separate datasets. Across these experiments, participants made decisions in a variety of tasks and reported how confident they were in their answers. The researchers used a hierarchical model designed to estimate fluctuations in each participant’s decision criterion on individual trials. Their preregistered prediction was that confidence would be systematically related to the criterion state active during the same trial. The prediction was supported in 14 of the 15 datasets, producing a remarkably consistent pattern across different experimental paradigms and methods for measuring confidence.

The result builds on signal detection theory, a framework widely used to explain how people distinguish meaningful signals from noise. In a basic signal detection model, a decision-maker converts incoming information into an internal decision variable. When that variable crosses a criterion, the person chooses one response rather than another. For example, someone deciding whether a faint sound contains a tone might compare the strength of the internal signal with a mental boundary. If the signal exceeds the boundary, they answer “present”; if it falls below it, they answer “absent.” Traditional accounts often assume that this criterion remains stable, at least over the short period covered by an experiment.

That assumption has increasingly come under pressure. People may adopt a more conservative criterion after a series of false alarms, become more liberal when signals are expected, or shift their response threshold as a result of attention, fatigue, motivation and arousal. These changes do not necessarily mean that the sensory evidence itself has altered. Rather, they change how much evidence is required before committing to a choice. The new study asks what such fluctuations mean for confidence. If confidence is calculated as the distance between the decision variable and the criterion, then a moving criterion should change confidence even when the underlying evidence remains exactly the same.

The computational logic is subtle but important. Imagine two trials in which the brain receives an identical internal signal. On one trial, the decision criterion is relatively low, making a “yes” response easy to produce. On another, the criterion is higher, requiring more evidence before the same response is selected. The distance between the signal and the criterion will differ across those trials, and so should confidence, even though the sensory input has not changed. Depending on the direction of the decision and the position of the criterion, a shift in the threshold can make a response feel more or less secure. Confidence, in this account, is not simply a readout of evidence; it is a readout of evidence interpreted through a changing decision boundary.

To test this possibility, the researchers applied a hierarchical model for fluctuations in criterion to trial-level behavioural data. Hierarchical modelling is especially useful when researchers need to estimate hidden psychological states that cannot be observed directly. Instead of calculating a separate criterion from a small number of responses for every individual trial, the model combines information across trials and participants while allowing the criterion to vary over time. This approach can distinguish relatively stable individual differences from short-term shifts in decision policy. The resulting trial-by-trial criterion estimates were then compared with participants’ confidence reports and with other signals linked to decision formation.

The pattern extended beyond subjective ratings. Criterion-state effects were also detected in reaction times, suggesting that momentary changes in the decision boundary influence how quickly people reach a response. The researchers additionally observed corresponding relationships with pupil-linked arousal, a physiological measure often associated with changes in alertness, effort and neuromodulatory state. A neural signature of confidence showed the same kind of relationship, indicating that the phenomenon is not confined to what participants say after making a decision. Together, these findings suggest that criterion fluctuations may be reflected across multiple levels of the decision system, from behaviour and physiology to brain activity.

The study’s breadth is central to its significance. Confidence experiments often use different stimuli, response formats and rating scales, making it difficult to know whether a result reflects a general property of cognition or a peculiarity of one task. Here, the effect appeared across 14 datasets involving a range of paradigms and confidence measures. That replication does not mean that every aspect of confidence is explained by criterion fluctuations, nor does it eliminate the role of sensory evidence. Instead, it indicates that the current state of the decision criterion is a robust contributor to confidence and may help explain why people can feel differently about decisions supported by comparable evidence.

The findings could reshape how scientists interpret confidence variability. In many studies, trial-to-trial changes in confidence are treated as measurement noise, inconsistent attention or unexplained idiosyncrasy. The new results suggest that at least some of this variability is informative. A person’s confidence may reveal not only what evidence was available, but also how demanding or permissive the decision system was at that particular moment. This perspective could prove relevant to research on eyewitness testimony, medical diagnosis, perceptual uncertainty and artificial intelligence, where confidence is often used to judge the reliability of a decision. It also offers a more dynamic picture of human judgment: the mind does not apply evidence to a permanently fixed rule, but continually adjusts the rule itself—and confidence tracks those adjustments in real time.