Participants who knew they were about to be given electric shocks were measurably calmer than participants who thought they might be. This is not a metaphor. It was confirmed in sweat glands and in pupil size. Cortisol, extracted from saliva, confirmed the paradigm was genuinely stressful overall \u2014 a real endocrine response, not just reported discomfort. The stress response \u2014 the actual, physiological machinery of alarm \u2014 was highest not when the outcome was certain and bad but when the outcome was unresolved. When the probability of a painful shock hovered near fifty percent, the body was more agitated than when the probability sat at one hundred.<\/p>\n\n\n\n
Fifty percent \u2014 in a laboratory paradigm, measured with biosensors, formalized in a computational model. Not “uncertain” in the loose, colloquial sense. Not anxious people speculating about the future. A precisely quantified midpoint of unresolved probability, and the point at which the human stress response peaks. Knowing you are going to be hurt was better for your nervous system, physiologically, than not knowing whether you might be.<\/p>\n\n\n\n
Most people’s first response to this is to assume they heard it wrong. The finding hits a pre-theoretical certainty that bad news is the thing that damages us \u2014 that dread tracks danger, that the stress response is a sensible barometer of threat severity. But the stress response is not a barometer of severity. It’s a barometer of resolution. And that distinction \u2014 quiet, almost technical-sounding \u2014 turns out to explain a great deal about the particular shape of suffering in the early twenty-first century.<\/p>\n\n\n\n
The evidence doesn’t point toward comfort. What follows is a reckoning with what it actually says.<\/p>\n\n\n\n
The study is Archy de Berker and colleagues, published in Nature Communications in 2016: “Computations of uncertainty mediate acute stress responses in humans.” Forty-five participants, a virtual rock-turning game, mild electric shocks to the hand. Painful enough to register as real stressors \u2014 not uncomfortable-tapping-on-the-wrist but actual aversive stimulation. The premise was simple: some rocks, when turned over, revealed a snake; a snake meant a shock. The probability of finding a snake under any given rock changed as the task progressed, and participants had to track it, updating their estimates through something like 320 trials.<\/p>\n\n\n\n
The design is doing real work here. The shocks weren’t assigned randomly. The probability mapping shifted continuously, which meant uncertainty wasn’t a static condition the researchers imposed \u2014 it was something that evolved moment by moment as participants learned the game. Sometimes you’d turned over enough rocks to know this one was safe; sometimes the probabilities were genuinely unresolved. The point is that certainty and uncertainty were both present, and measurably different, within each participant across the session.<\/p>\n\n\n\n
Skin conductance and pupil dilation tracked what happened to the body as the probabilities shifted. Both measures \u2014 physiological indicators of arousal that bypass any deliberate attempt at self-presentation \u2014 rose and fell with the uncertainty estimates, not with the proximity to a guaranteed shock. Subjective stress ratings did the same. When participants estimated their shock probability at around fifty percent, all three measures peaked. At zero percent and at one hundred percent, they converged. The body treated “definitely getting shocked” and “definitely not getting shocked” as physiologically equivalent. “Might get shocked” was the problem.<\/p>\n\n\n\n
Cortisol \u2014 measured from saliva \u2014 confirmed that the experiment was genuinely stressful overall. The researchers were careful about what the cortisol data could support: it establishes that the paradigm induced real physiological stress, which it did. It’s the skin conductance and pupil data that captured the uncertainty-specific dynamics, and those are the findings that matter for the argument.<\/p>\n\n\n\n
Worth noting what the paradigm found about learning, too. The uncertainty tuning of stress predicted individual task performance \u2014 participants whose stress responses were appropriately calibrated to uncertainty were the ones who learned the task better. The stress was functional. A feature, not a malfunction, and one that matters for what follows.<\/p>\n\n\n\n
\n\nThe machine that measured the uncertainty<\/strong>\n\nThe researchers didn't just track what participants reported feeling. They used a hierarchical Bayesian computational model \u2014 borrowed directly from the mathematical framework used to understand how brains update beliefs under uncertainty \u2014 to reconstruct each participant's running probability estimate from their trial-by-trial behavior. What the model did was infer, at each moment in the task, what subjective probability distribution the participant was operating with. Not what they said they thought, but what their behavior implied they thought.\n\nThis matters because it rules out simpler stories. The stress response wasn't correlated with how worried participants said they were. It was correlated with formal estimates of subjective uncertainty extracted from behavioral data. That's a different claim. It means what de Berker and colleagues were measuring is not \"people who feel anxious report being stressed\" but something more specific: the uncertainty being computed by the brain's prediction machinery was driving physiological responses independent of what participants consciously stated. The formal model connected the internal computation to the bodily output. The two tracked each other, precisely, across hundreds of trials.<\/code><\/pre>\n<\/div><\/div>\n<\/blockquote>\n\n\n\nWhat the brain is actually doing<\/h3>\n\n\n\n
The stress system doesn’t care how bad the outcome is. It activates in proportion to how unresolved the prediction is. These sound like they should be the same thing. They’re not.<\/p>\n\n\n\n
The anterior cingulate cortex has been one of the most studied regions in the neuroscience of decision-making and control for the past several decades \u2014 and also one of the most debated. Its precise function is contested. Conflict monitoring, error detection, attention allocation, performance evaluation: different researchers have emphasized different aspects, and none of these characterizations is obviously wrong. What they converge on is a region acutely sensitive to the gap between what was expected and what happened. Prediction error, in various formulations, is the ACC’s core business.<\/p>\n\n\n\n
Alexander and Brown’s 2019 paper in Topics in Cognitive Science develops this through the PRO model \u2014 Predicted Response-Outcome \u2014 in which the ACC attempts to predict all potential outcomes and signals surprise when events deviate from what was expected, positive or negative. Unlike standard reinforcement learning frameworks where positive and negative prediction errors track good and bad outcomes, the PRO model’s signals are about unexpectedness \u2014 events that occurred but weren’t predicted, and events that were predicted but didn’t occur. Shenhav, Botvinick, and Cohen’s 2013 Neuron paper extends the framework to cognitive control, arguing that the ACC allocates attention and effort based on expected value of control \u2014 which itself depends on how clearly the environment’s structure can be read. Unresolved probability is high-information-demand territory.<\/p>\n\n\n\n
The implication is precise. When the brain has resolved a prediction \u2014 even a bad one \u2014 it has an answer. The computation terminates, or at least reaches a stopping condition. What’s coming can be braced for; the motor and endocrine systems can prepare. Certain-bad is a problem, but it’s a solved problem. The brain knows the state of the world and can respond accordingly.<\/p>\n\n\n\n
Uncertain-anything is unsolved. And unsolved problems keep the computation running.<\/p>\n\n\n\n
Not dread about a future outcome but the subjective texture of a system computing continuously without arriving at resolution. The ACC’s prediction-error machinery returns a probability distribution that won’t collapse, and it keeps returning it, and the stress response that accompanies sustained high-uncertainty computation is not about imagining how bad the shock will be. It’s the running of the calculation. The discomfort is the computation, not its object.<\/p>\n\n\n\n
The content of the uncertain outcome \u2014 the thing you might be dreading \u2014 is less relevant to the stress response than the degree to which the probability is unresolved. Change the stakes but keep the probability distribution and the physiological signature looks similar. Keep the stakes constant but resolve the probability and the signature drops. Which is exactly what de Berker and colleagues found.<\/p>\n\n\n\n
Why the brain was built this way<\/h3>\n\n\n\n
A quick note on evolutionary framing: what follows is an inference from known mechanisms, not a settled empirical account of how this particular neural architecture was selected for. Evolution doesn’t leave lab notes.<\/p>\n\n\n\n
That said, Randolph Nesse’s smoke detector principle, developed in a 2001 paper in the Annals of the New York Academy of Sciences, provides the relevant logic. When the cost of a false alarm is low and the cost of a miss is catastrophic, a well-calibrated system will produce many false alarms. Anxiety, Nesse argues, is a system that’s been calibrated precisely this way \u2014 not to minimize subjective distress but to minimize the probability of a fatal miss. The frequent minor suffering from false alarms is not a design flaw. It’s the cost of the design working as intended.<\/p>\n\n\n\n
Apply this to uncertainty specifically: in environments where threat was physical and predators were real, the animal that became maximally vigilant upon detecting any evidence of possible danger \u2014 not confirmed danger, possible danger \u2014 was the one more likely to survive. Waiting for visual confirmation before initiating the stress response was a bad strategy. The stress response is supposed to fire early, on incomplete information. “Something might get you” is the trigger. That’s the feature.<\/p>\n\n\n\n
Which means the brain’s preference for resolved-bad-outcome over unresolved-possible-outcome has a straightforward logic. Confirmed threats could be acted on. Uncertain threats required sustained surveillance \u2014 expensive, but in ancestral environments this was affordable, because uncertainty was almost always bounded. The predator either got you or it didn’t. Seasons ended. Illness resolved one way or the other. The probability distributions reached zero or one, and the computation could stop. The surveillance period had a natural endpoint.<\/p>\n\n\n\n
The problem is not that the system is irrational. The problem is what happens when resolution is withheld indefinitely.<\/p>\n\n\n\n
The specific way modern environments break this<\/h3>\n\n\n\n
Not the quantity of uncertainty but its resolvability. That’s the distinction the evolutionary argument was pointing toward.<\/p>\n\n\n\n
Consider two cases, same mechanism, different duration. You’ve had a biopsy and are waiting for results. The uncertainty is real, the stakes are real, the stress response is running. But the endpoint is defined: the test is processed, the results come in, the probability collapses to zero or one. The computation terminates. This is an acutely stressful experience that the brain is, on some level, built to manage \u2014 because the structure is familiar: uncertainty with a natural resolution.<\/p>\n\n\n\n
Now consider precarious employment. A gig contract, a performance assessment with no defined criteria, an “at will” arrangement that depends on consumer ratings that can change week to week. The uncertainty is not episodic. There is no test result, no review date on the calendar, no moment at which the probability resolves. The surveillance mechanism runs on standby indefinitely. The brain’s prediction machinery keeps computing because there is never an output that says: okay, done, you either have the thing or you don’t.<\/p>\n\n\n\n
Sverke, Hellgren, and N\u00e4swall’s 2002 meta-analysis in the Journal of Occupational Health Psychology documented the consequences of this over hundreds of studies: job insecurity has detrimental effects on health, job attitudes, and organizational outcomes, and these effects are robust across contexts.<\/p>\n\n\n\n
The IU construct \u2014 intolerance of uncertainty \u2014 was developed to characterize precisely this: the degree to which a person finds any unresolved probability aversive, independent of what the probability is about. High IU does not mean overestimating the likelihood of bad outcomes. It means finding it intolerable that the outcome is unresolved at all. The patient who is told their scan results show a ninety percent chance of being fine does not experience ten percent of the distress of someone told there was a certainty of a problem. They experience something much closer to the full distress, because there’s still ten percent. Still unresolved. Still running.<\/p>\n\n\n\n
IU, developed by Mark Freeston, Michel Dugas, and colleagues, is measured by the Intolerance of Uncertainty Scale \u2014 twenty-seven items, internal consistency alpha of 0.91, two-factor structure. What it notably does not ask about is content. No specific feared outcomes. No probability estimates for particular bad events. Just the relationship to not-knowing as a general state.<\/p>\n\n\n\n
\n\nWhat the Intolerance of Uncertainty Scale actually measures<\/strong>\n\nThe IUS doesn't ask what you're worried about. It asks things like: \"Unforeseen events upset me greatly.\" \"Not knowing what will happen makes me unable to function properly.\" \"I can't stand being undecided about my future.\" \"When I am uncertain, I can't function very well.\" No reference to any specific feared content \u2014 no cancer, no financial ruin, no relationship failure. Just: how do you relate to the state of not-knowing?\n\nThe scale's two-factor structure reflects two ways that uncertainty-intolerance manifests. The prospective factor captures people who experience any uncertainty as an active threat \u2014 anxiety about uncertainty as a state of being. The inhibitory factor captures people who freeze: uncertainty doesn't just feel bad, it prevents action. Both factors predict anxiety and worry; both are transdiagnostic. What makes the IUS striking as a clinical instrument is that it predicts GAD, OCD, health anxiety, depression, and features of PTSD better than most disorder-specific measures \u2014 because it's measuring the underlying variable, not the domain-specific manifestation of it. A system that treats unresolved probability as inherently threatening will locate threatening content wherever it turns its attention.<\/code><\/pre>\n<\/div><\/div>\n<\/blockquote>\n\n\n\nIU predicts across disorders, not within them \u2014 GAD, OCD, health anxiety, depression, features of PTSD \u2014 because it’s measuring the underlying variable, not the domain-specific expression of it. The specific worry domain is where the problem lands, not what it is. Same intolerance, different address. Hebert and Dugas, writing in 2019 in Cognitive and Behavioural Practice, document that approximately twenty to thirty percent of GAD patients don’t achieve full remission after standard CBT; in the non-responders, IU remains elevated. These patients are not failing because their feared outcomes are more severe. They’re failing because the underlying variable \u2014 the intolerance of uncertainty itself \u2014 hasn’t moved.<\/p>\n\n\n\n
What the misfire implies about what actually helps<\/h3>\n\n\n\n
Most anxiety management doesn’t target this. That’s the uncomfortable implication.<\/p>\n\n\n\n
Cognitive restructuring asks whether the feared outcome is realistic. Is the catastrophe you’re imagining actually likely? That’s a reasonable question, but it’s aimed at the content \u2014 the specific probability estimate \u2014 not at the intolerance of the unresolved state. Reassurance-seeking tries to push the probability closer to zero. Thought suppression removes uncertain content from attention. All three approaches are trying to make the probability feel more resolved. They work, briefly. Then a new uncertainty arises, and the loop restarts.<\/p>\n\n\n\n
Reassurance-seeking is the instructive case. The person who repeatedly asks their doctor whether a symptom is serious is not irrational. They’re doing exactly what their brain is asking for: trying to resolve the probability estimate. The problem is that reassurance-seeking trains the expectation that resolution is always available. Every new uncertainty becomes a trigger for the seeking behavior, because the last one was resolved by asking. The brain habituates not to tolerating ambiguity but to expecting resolution. IU increases. It’s a well-documented pattern in GAD and health anxiety both.<\/p>\n\n\n\n
The alternative \u2014 IU-targeted CBT, developed over several decades primarily by the Dugas group \u2014 is not exposure to feared outcomes but exposure to the state of not-knowing. Behavioral experiments where the patient deliberately does not seek resolution, and observes that the uncertainty is survivable. The prediction machinery runs; it doesn’t terminate; and nothing catastrophic follows. Over repeated trials, the relationship to unresolved probability shifts. Not: certainty is achieved. But: unresolved probability is tolerable.<\/p>\n\n\n\n
Bomyea and colleagues, in a 2015 paper in the Journal of Anxiety Disorders, examined this mechanism in a CBT sample with GAD. Reductions in IU mediated subsequent reductions in worry, and those reductions in IU accounted for fifty-nine percent of the worry reduction observed over the course of treatment. More than half of the therapeutic benefit came through the IU channel. Standard CBT produces IU reductions \u2014 but largely, the data suggests, as a byproduct of behavioral activation and behavioral experiments that incidentally build tolerance for unresolved probability. The treatments work partly for reasons the field hasn’t always named.<\/p>\n\n\n\n
The meta-analytic view, from Miller and McGuire’s 2023 paper in the Journal of Affective Disorders examining twenty-eight randomized controlled trials, found a controlled effect size of Hedges’ g = 0.89 for IU improvement \u2014 large by conventional standards. Treatment effects on IU predicted improvements in symptom severity and accounted for thirty-six percent of the variance in outcomes.<\/p>\n\n\n\n
The relevant therapeutic move is something more specific than reducing worry. It’s distinguishing uncertainties that can be resolved from those that structurally cannot. The biopsy waiter can confirm: the test is being processed, the timeline is X, a specific person will call. They cannot confirm the result. Conflating these two \u2014 treating the result as something that could be known if only they could find the right person to ask \u2014 keeps the uncertainty system fully loaded across both. Separating them gives the prediction machinery a termination condition for the resolvable subset. It’s not a solution; it’s the closest approximation to the relief the brain is actually seeking.<\/p>\n\n\n\n
What “works” does not mean eliminating discomfort. Not achieving certainty. It means shifting the relationship to unresolved probability from intolerable to uncomfortable-but-survivable. That’s a more modest goal than most popular anxiety discourse offers. It’s also more achievable, and more honest about what’s actually happening.<\/p>\n\n\n\n
The architecture of permanent unresolvability<\/h3>\n\n\n\n
The evolutionary argument was about bounded uncertainty. The predator either got you or it didn’t. Modern architecture isn’t like that. And in some cases, this isn’t an accident.<\/p>\n\n\n\n
Social media platforms deliver social evaluation signals on a schedule that is, deliberately or incidentally, irregular. Likes arrive unpredictably. Comments may or may not materialize. A post goes out into a void and might resonate or might not, and there’s no timeline on finding out, and there will always be another post after this one. This maps structurally onto B.F. Skinner’s variable ratio reinforcement schedule \u2014 the most behaviorally robust pattern of conditioning, most resistant to extinction \u2014 and by the logic of the de Berker mechanism, that arrangement is exactly what would keep the uncertainty-computation running at low but continuous load. The probability never reaches zero or one. The surveillance mechanism has no natural off switch. Vannucci and colleagues, in a 2017 study in the Journal of Affective Disorders, found that increased social media use correlated with higher anxiety among emerging adults \u2014 an association that’s correlational, but the mechanism the lab work predicts is exactly this: a source of social evaluation that is constitutively unresolvable.<\/p>\n\n\n\n
Precarious employment works the same way. No contract security, no defined performance criteria, no review date. Consumer ratings on platform apps that fluctuate with factors outside the worker’s control. Sverke and colleagues documented the health effects of job insecurity in 2002; the gig economy has since refined the mechanism \u2014 embedding structural uncertainty into the terms of employment rather than just the labor market. Income is uncertain. Status is uncertain. Whether you’ll have work next month is uncertain. None of these resolve at a defined point. The brain’s uncertainty computation is running on open loop.<\/p>\n\n\n\n
News media sustains possible threats in the worst state for the brain’s uncertainty machinery: unresolved, kept open rather than allowed to close. The story is developing. The investigation is ongoing. The threat is real but unconfirmed. It is worth noting that this structural observation carries less evidential weight than the social media and employment claims \u2014 the causal chain from news coverage to anxiety is more complex and contested than either \u2014 but as a structural observation about how threats are framed and maintained, the pattern holds.<\/p>\n\n\n\n
The convergence across these environments isn’t, in most cases, the result of deliberate design. Nobody sat in a room and decided to engineer an environment that would keep the human uncertainty-computation running continuously at high load. Platforms optimized for engagement found, empirically, that unpredictable reward schedules generate more checking behavior. Employment structures evolved to minimize employer risk and maximize flexibility. News cycles discovered that unresolved threat is more compelling than resolved threat. The result is environments that happen to map, with remarkable precision, onto exactly the conditions that the 2016 de Berker paradigm measured as most physiologically costly.<\/p>\n\n\n\n
The participants in that study eventually left the lab. They turned the last rock, the session ended, the probability estimate closed. The snake was there or it wasn’t. They got their shock, or they didn’t, and then they went home.<\/p>\n\n\n\n
Modern life’s architecture does not, in general, provide that termination condition. The rocks keep appearing. The snakes remain possible. And the uncertainty-computation \u2014 built for episodic threat in bounded environments \u2014 continues running, on hardware that was never designed for the task.<\/p>\n\n\n\n