Every decision carries an implicit question: How likely is the choice to be correct? Confidence—the answer to that question—tells a decision-maker how much to trust their own choice. Confidence shapes how we communicate our beliefs, learn from feedback and decide whether to act or seek more information.
For a long time, confidence occupied what psychologist Douglas Vickers called “a Cinderella role in cognitive psychology—relied on for its usefulness but overlooked as an interesting variable in its own right.” That has changed. Researchers now ask where confidence is represented in the brain, how it is computed, how it breaks down in psychiatric disorders and even how to endow artificial agents with basic metacognitive capacities. One question sits at the center of this work: What exactly is the brain computing when it reports confidence?
The prevailing view is that choice and confidence are organized hierarchically: The brain first makes a decision—even if only provisionally—and then estimates how likely that decision is to be correct. In this view, confidence is constructed by reading out features of the process that produced the choice: how much evidence accumulated for each option, how long it took to decide, how decisively the winning option outpaced the alternatives or how many times the decision-maker changed their mind during deliberation. This framework has been extremely successful in explaining choice, response time and confidence judgments in a wide variety of tasks and has secured a powerful idea: Confidence is a derived quantity, read out from the decision process.
The neurobiology of decision-making, however, invites another, less intuitive, possibility. When each option is mapped onto a particular action in advance, the decision unfolds in the circuits that prepare those actions. Consider a monkey deciding whether a cloud of moving dots has net leftward or rightward motion. If the monkey reports its choice with an eye movement, neurons involved in planning eye movements—including a subset of those in the lateral intraparietal area—represent the evolving decision variable. Their activity does not primarily encode an abstract judgment such as “the motion is rightward.” Instead, it reflects the growing intention to look toward the rightward target. In this view, deciding is inseparable from preparing to act, the essence of what Michael Shadlen and his colleagues called an “intentional” framework.
T
his framework suggests a different way to think about confidence. If a decision can unfold as a competition among actions rather than abstract propositions, confidence can be incorporated into that same competition. Imagine a decision-maker reporting both perceptual choice and confidence with a single eye movement by choosing among four targets: left-high confidence, left-low confidence, right-low confidence and right-high confidence. A hierarchical account describes this as a sequence of computations: Determine the direction, evaluate confidence and then select the appropriate target.But the intentional framework permits a different interpretation: The four actions may themselves be the alternatives over which the decision unfolds. Different states of sensory evidence favor different joint choice-confidence responses: Strong rightward evidence, for example, favors the right-high-confidence action, whereas weak rightward evidence may favor the right-low-confidence action (as in the RTCON model). In this view, sensory evidence feeds four parallel accumulators in eye-movement-planning circuits, one for each possible response. The first accumulator to reach a threshold determines both choice and confidence.
The hierarchical and intentional accounts make different predictions about the role of action-preparation circuits. In a hierarchical account, choice and confidence can be resolved upstream and then translated into the action used to report them. Even if confidence is read out continuously from an evolving decision, activity in action-preparation circuits may reflect the evolving decision rather than participate in the decision process itself. In an intentional account, by contrast, the decision unfolds within these circuits: Signals related to choice and confidence should emerge together as the brain selects among the joint choice-confidence actions.
Although no neurophysiological study of confidence has directly compared hierarchical and intentional accounts, a 2025 study provides some clues. The study implemented a version of the four-target task described above: Macaques reported both motion direction and confidence by selecting one of four eye-movement targets while the researchers recorded from lateral intraparietal neurons representing individual targets. Activity distinguishing high- from low-confidence targets began to diverge at approximately the same time as activity distinguishing leftward from rightward choices, roughly 200 milliseconds after motion onset. Confidence-related activity, in other words, did not appear to lag behind choice-related activity. These findings are difficult to reconcile with a strictly hierarchical account in which confidence evaluation begins only after commitment to a choice, and they are naturally accommodated by an intentional account, in which both dimensions of the report emerge together within an eye-movement-planning circuit.
This does not imply that confidence must always be built into the decision itself. In many tasks, confidence dissociates from choice in ways that seem to require a distinct evaluative process. My own recent work suggests that confidence may be implemented intentionally or hierarchically depending on the context. An intentional account appears to explain behavior better when confidence reports are explicitly incentivized—for example, through a point system that rewards accurate use of the confidence scale—whereas a hierarchical strategy is favored when confidence judgments carry no explicit consequences. Confidence may therefore not have a single computational architecture. Sometimes it behaves like a genuine second thought; sometimes it is built into the first.
Taking the intentional view seriously changes the experimental questions we should ask. How is sensory evidence categorized and routed into actions that jointly represent choice and confidence? Is confidence itself shaped by the actions available to report it—for example, whether the same belief is communicated verbally, numerically, via gaze or by pointing? And what determines whether the brain adopts an intentional or hierarchical architecture?
Confidence has come a long way from its Cinderella role. The emerging debate is no longer about whether confidence matters but about what kind of computation it is. Perhaps confidence is not always the brain’s second thought about a decision. Sometimes it is simply the outcome of a richer one.
