Sometimes the most informative answer is not yes or no.
It is: "I do not have enough evidence yet."
Uncertainty monitoring concerns the detection of insufficient or unreliable information about one's own likely performance and the use of that uncertainty in behavior. It can appear as lower confidence, information seeking, delayed commitment, checking, or an option to decline a difficult trial.
The construct sits inside the broader territory of metacognitive monitoring, but it has an independent research history, including comparative cognition, where researchers have asked whether nonhuman animals respond adaptively to uncertainty.
Its scientific value lies in the boundary it forces us to respect: behavior can indicate sensitivity to uncertainty without proving a human-like conscious experience of "knowing that I do not know."
The uncertain response made the phenomenon experimentally visible
Comparative studies often give subjects a difficult discrimination task together with an "uncertain" or opt-out response. If that response is used selectively on the hardest trials, it suggests that the subject is sensitive to conditions under which error is more likely.
J. David Smith and colleagues developed a substantial comparative literature using such paradigms with humans and other animals. Monkeys and dolphins, among other species studied in this tradition, sometimes use uncertainty responses adaptively rather than randomly.
The finding is important because it demonstrates flexible behavior around uncertainty.
The interpretation is harder.
An animal may decline a difficult trial because of a metacognitive representation of uncertainty, because learned stimulus-response associations track difficulty, or because other mechanisms produce similar behavior. Experimental design is therefore not merely procedural. It determines which explanations remain plausible.
Functional uncertainty and conscious uncertainty are different claims
In ordinary human language, "I am uncertain" implies a subjective state.
Scientific evidence does not allow that meaning to be imported automatically into every uncertainty-sensitive behavior.
Comparative cognition has long debated whether opt-out responses demonstrate metacognition or can be explained by lower-level learning mechanisms. Reviews such as Smith, Shields and Washburn's 2003 Behavioral and Brain Sciences article argue for important functional parallels. Later work, including Jonathon Crystal's 2019 perspective, emphasizes both the promise and unresolved methodological problems of comparative metacognition.
The careful position is layered.
Evidence can support uncertainty-sensitive behavior. Stronger claims about reflective awareness require stronger evidence. Claims about phenomenology may remain inaccessible.
That boundary is not a weakness. It is what makes the field intellectually honest.
Human uncertainty monitoring is broader than opting out
People manage uncertainty through many actions.
A reader notices they cannot explain a paragraph. A doctor orders another test. A researcher withholds a conclusion because the confidence interval is wide. A driver slows when visibility drops. A student marks an answer for review.
Some actions are metacognitive because they are guided by information about one's own likely correctness or knowledge. Others are direct responses to environmental uncertainty.
The distinction matters.
Not every uncertain environment requires metacognition. A storm warning can directly change behavior. Metacognitive uncertainty specifically concerns uncertainty about one's own cognitive state, decision or performance.
Confidence and uncertainty are related but not identical
Low confidence often indicates high uncertainty, but the two terms are not perfect inverses in every model or task.
Fleming's review distinguishes personal-level confidence judgments from subpersonal uncertainty representations. A sensory system can represent noisy evidence without the person forming an explicit belief about their own correctness.
At the personal level, uncertainty can be expressed as low confidence, hesitation, request for more information or refusal to answer.
At the subpersonal level, uncertainty may affect processing before any report exists.
This layered architecture helps prevent an increasingly common error in AI discourse, where any probability distribution, confidence score or hesitant phrase is described as if a system "knows that it does not know."
Those are different claims.
Uncertainty can trigger control
Monitoring matters when it changes action.
If low confidence leads to a second check, information seeking or deferred commitment, uncertainty becomes part of a monitoring-control loop. If the person recognizes uncertainty but is forced to answer, the signal exists without regulatory freedom.
This is why an opt-out option is theoretically interesting. It gives uncertainty somewhere to go.
In human institutions, equivalent options include escalation, consultation and the ability to say "insufficient evidence." A culture that requires immediate certainty can suppress useful control even when individuals privately detect uncertainty.
The Institute's interest in uncertainty-tolerant systems is an application synthesis. The established science concerns monitoring, confidence, information seeking and decision behavior. Whether a particular institutional design improves epistemic quality needs direct evaluation.
Measurement is vulnerable to alternative explanations
Suppose participants use an uncertain option mainly on difficult trials. Several mechanisms may fit the pattern.
They might monitor an internal confidence signal. They might learn that certain stimulus features predict low reward. They might respond to reaction-time cues. They might use a simple threshold on perceptual evidence without representing their own knowledge.
Researchers therefore use transfer tests, novel stimuli, response competition, task changes and other designs intended to distinguish mechanisms.
No single task settles the philosophical question.
This is one reason uncertainty monitoring deserves an independent page rather than being compressed into a paragraph under confidence. Its literature is partly defined by the contest between functional evidence and interpretive overreach.
Uncertainty has an economic structure
Choosing uncertainty can carry costs.
An opt-out response may yield a smaller reward than a correct answer but a better outcome than an error. Seeking more information consumes time. Deferring a decision may lose an opportunity. Asking for a second opinion can carry social cost.
The threshold for acting on uncertainty therefore depends on payoff structure as well as subjective confidence.
This has direct relevance to real-world decision systems. In a low-cost setting, checking may be easy. In emergency medicine, aviation or cybersecurity, delay can itself be dangerous.
The scientific question is not whether uncertainty should always produce caution. It is how monitoring interacts with consequences and available actions.
Intellectual humility is not the same construct
Intellectual humility concerns recognizing the fallibility or limitations of one's beliefs and knowledge, often with dispositional and interpersonal dimensions.
Uncertainty monitoring can be much narrower and more local.
A person may be excellent at detecting low confidence on perceptual tasks while being dogmatic in social beliefs. Another may endorse intellectual humility yet show poor calibration on specific decisions.
The constructs may be related in some contexts, but evidence for one cannot be treated as evidence for the other.
This distinction protects both pages from becoming generic celebrations of doubt.
The AI problem is uncertainty translation
AI systems can expose numerical uncertainty, produce caveats, retrieve sources or generate fluent answers with no visible hesitation. Human users then have to translate those signals into their own monitoring and control.
There is no single uncertainty channel.
A model's token probability is not automatically a calibrated probability that a factual claim is correct. A verbal hedge may be stylistic. A citation can exist and still fail to support the sentence. The user may be highly confident because the prose is fluent even when system-level uncertainty is high.
Research on these human-AI interactions is developing rapidly. The Institute should use classic uncertainty-monitoring research to structure questions, not to claim that the problem is already solved.
The distinction between uncertainty and ambiguity matters
Decision science often distinguishes risk, where probabilities are known or estimable, from ambiguity, where the probability structure itself is unclear. Metacognitive uncertainty adds yet another layer: uncertainty about one's own judgment or knowledge.
These can coexist.
A person may face an ambiguous external situation and be highly confident in a bad model of it. Another may face a well-characterized probability distribution but be unsure they have interpreted it correctly.
The distinction prevents uncertainty monitoring from becoming a catch-all label for every difficult decision.
Information seeking is one of the clearest behavioral consequences
When people recognize that current evidence is insufficient, one adaptive response is to seek more information.
That action can be studied directly: do participants request another sample, inspect a source, consult an expert, or defer until evidence improves?
Information seeking also reveals costs. If another sample is expensive, uncertainty must cross a higher threshold before search is worthwhile. If information is cheap, repeated checking can become excessive.
For AI interfaces, this suggests a measurable design question: does the system help users seek independent evidence when uncertainty is high, or merely encourage another generation from the same model?
The latter may feel like verification while adding little independence.
Uncertainty tolerance is not identical to uncertainty monitoring
Psychological literature also uses terms such as intolerance of uncertainty to describe emotional or dispositional responses to uncertain situations.
Those constructs are related to how people experience uncertainty, but they are not the same as detecting when a particular cognitive judgment is unreliable.
A person can dislike uncertainty intensely while monitoring it accurately. Another can be comfortable with ambiguity and still be poorly calibrated.
Keeping these constructs separate is important for both research and public communication. Otherwise, a precise metacognitive mechanism quietly turns into a broad personality story.
Institutions can hide uncertainty by aggregation
When many estimates are combined into a report, the final number can look more certain than the underlying evidence.
Ranges disappear. Dissent is summarized away. Assumptions become footnotes. The presentation layer converts a distribution of uncertainty into one clean recommendation.
An institutional uncertainty architecture would preserve enough provenance to show where uncertainty entered, how it was combined and which decisions remain sensitive to it.
That is an Institute design synthesis, not an established psychological construct. But it is directly motivated by the research lesson that uncertainty only supports control when it remains represented rather than erased.
Uncertainty can be informative before it becomes verbal
People sometimes slow down, recheck or seek additional evidence before they can explain what feels wrong. That does not automatically establish a distinct unconscious metacognitive system, but it reminds us that explicit verbal report is not the only observable consequence of uncertainty.
Reaction time, information sampling and choice patterns can all provide evidence about how uncertainty affects behavior. Each measure has alternative explanations, so converging evidence is stronger than any one signal.
This matters for populations or species where language cannot carry the burden of proof.
There is a difference between preserving uncertainty and amplifying it
An institution that hides uncertainty creates overconfidence. An institution that foregrounds every possible doubt can create paralysis.
Useful uncertainty communication needs prioritisation. Which unknowns could change the decision? Which are merely residual? Which can be reduced with additional evidence? Which are irreducible within the available time?
These questions move beyond laboratory uncertainty monitoring into decision architecture. They are Institute synthesis, but they preserve the core scientific insight: uncertainty has value when it is discriminating enough to guide action, not when it simply increases the amount of doubt in the room.
What the evidence does not show
Uncertainty-sensitive behavior does not by itself prove conscious metacognition.
Using an opt-out response on difficult trials does not uniquely identify one cognitive mechanism.
Low confidence is not always equivalent to a rich reflective state of "knowing that one does not know."
Evidence from animal uncertainty tasks does not justify claims about subjective animal experience without additional argument.
Uncertainty monitoring is not the same as intellectual humility, anxiety, indecision or generalized risk aversion.
More uncertainty is not always better, and more checking is not always adaptive.
And classic uncertainty-monitoring research does not establish that an AI system possessing probability estimates has human-like self-knowledge.
Open questions
Which paradigms best distinguish metacognitive monitoring from learned responses to task difficulty? How much uncertainty monitoring generalises across perception, memory and reasoning? When do people convert uncertainty into information seeking, and when do incentives suppress that conversion? Which uncertainty displays improve human calibration rather than merely adding another cue?
Comparative work faces a particularly hard problem: how can experiments provide evidence for flexible second-order control without assuming language or introspective report?
Human-AI research adds another: how should uncertainty move between machine representations and human decisions without being mistranslated into either false reassurance or unusable caution?
The right to say "not enough"
Uncertainty monitoring is often discussed as an internal capacity, but its consequences depend on whether a system permits an uncertain response.
A learner can know they do not understand and still be forced to submit an answer. A professional can detect weak evidence and still be rewarded for certainty. An algorithm can carry uncertainty information that an interface hides.
At the individual level, uncertainty can guide checking and learning.
At the institutional level, procedures decide whether uncertainty can change action.
At the civilizational level, increasingly powerful systems make confident output cheap, which may make disciplined expressions of "not enough evidence" more valuable.
The science does not tell us to glorify doubt.
It shows why adaptive cognition needs a way to represent the possibility of being wrong.
Research references
- Smith JD, Shields WE, Washburn DA. The comparative psychology of uncertainty monitoring and metacognition. Behavioral and Brain Sciences. 2003;26(3):317-339; discussion 340-373. https://doi.org/10.1017/S0140525X03000086
- Smith JD, Washburn DA. Uncertainty Monitoring and Metacognition by Animals. Current Directions in Psychological Science. 2005;14(1):19-24. https://doi.org/10.1111/j.0963-7214.2005.00327.x
- Crystal JD. Comparative Approaches to Metacognition: Prospects, Problems, and the Future. Animal Behavior and Cognition. 2019;6(4):254-261. https://doi.org/10.26451/abc.06.04.05.2019
- Fleming SM. Metacognition and Confidence: A Review and Synthesis. Annual Review of Psychology. 2024;75:241-268. https://doi.org/10.1146/annurev-psych-022423-032425
- Fleming SM, Lau HC. How to measure metacognition. Frontiers in Human Neuroscience. 2014;8:443. https://doi.org/10.3389/fnhum.2014.00443
- Rahnev D. A comprehensive assessment of current methods for measuring metacognition. Nature Communications. 2025;16:701. https://doi.org/10.1038/s41467-025-56117-0