How Do We Live with Uncertainty?

Uncertainty is not a flaw in the system

In a weather forecast, the question is not only:

will it rain or not?

There is also:

what is the probability of rain?

The National Academies emphasizes that uncertainty is a fundamental feature of forecasting and that a forecast is incomplete without an appropriate description of uncertainty.[1] This applies much more broadly. Uncertainty exists in:

• medical diagnosis;
• economic forecasting;
• historical attribution;
• technical risk assessment;
• legal proceedings;
• scientific models.

So the following ideal is mistaken:

good information = information without uncertainty.

Often the opposite is true:

good information also includes an honest description of what we do not know.

Risk and uncertainty are not always the same thing. Decision theory often distinguishes between a situation in which:

we know or estimate the probabilities of outcomes. and a situation in which:

even the probabilities themselves are not well determined. The Stanford Encyclopedia's discussion of decision theory describes standard expected-utility models and notes that real decisions often involve both subjective probabilities and deeper forms of uncertainty.[2] A simple example:

with a fair coin toss, we approximately know the probability.

A harder example:

the probability of a major technological breakthrough within the next ten years.

There is no stable frequency from which we can simply calculate an answer.

Sometimes we do not even know all possible outcomes. The deepest problem is not:

we do not know which of five outcomes will occur.

It can be:

we do not know whether we have even listed all important outcomes.

This happens with:

• new technologies;
• social crises;
• complex systems;
• rare catastrophes.

It is therefore useful to distinguish:

quantified risk. from

broader or structural uncertainty. The IPCC methodology for communicating uncertainty, for example, distinguishes among situations ranging from those in which we know only the direction of possible change to those in which we can estimate a range, a likelihood, or even a full probability distribution.[3]

“I don't know” has several meanings. The sentence:

“I don't know.”

can mean:

I do not have the information.

The information exists, but I have not checked it yet.

The data are poor.

Experts disagree.

The event is probabilistic by nature.

The model has a wide uncertainty range.

The question cannot currently be resolved empirically.

These are very different situations. So a better question is:

Why don't we know?

Uncertainty does not mean all options are equally likely. If we say:

“We are not 100 percent certain,”

it does not follow that:

50 : 50.

For example: meteorologists may estimate an 80% chance of rain. Uncertainty exists. But rain and dry weather do not have equal evidentiary support. This is one of the most common problems in public debate:

the presence of uncertainty is presented as the absence of knowledge.

Even 99% is not 100%. The opposite error is:

very high probability = certainty.

If we estimate something at 99% probability, we must be prepared for some outcomes in a suitably large class of similar cases not to occur. A good probabilistic forecaster is not someone whose 90% forecasts all come true.

If every one of them came true, the estimates might even be underconfident.

Calibration: do our confidence levels mean what we think they mean?

Imagine one hundred claims about which we said:

“I am 90% confident.”

If we are well calibrated, we would expect roughly:

90 correct

and

10 incorrect.

This is calibration. In probabilistic forecasting, calibration is a central criterion of forecast quality.[4] It changes the question. Instead of asking:

“Was I right?”

we ask:

“Over time, do my stated confidence levels match actual outcomes?”

One hit tells us almost nothing about calibration. If someone says:

“I predicted the event and it happened.”

that does not mean:

I am a good forecaster.

Chance also produces isolated hits. To judge forecasting quality, we need:

• many forecasts;
• probabilities recorded in advance;
• a clear outcome criterion.

This closely resembles scientific replication. The process matters more than one spectacular hit.

Brier score: a forecast pays for confidence. In 1950, Glenn Brier proposed a statistical method for evaluating probabilistic forecasts.[5] The Brier score rewards forecasts that:

• assign high probability to events that happen;

and penalizes:

• confident forecasts that turn out to be wrong.

The important idea is not the formula itself. It is the discipline:

if I say 95%, I accept a greater evidentiary responsibility than if I say 60%.

Two people can be equally often “right” and still not be equally good forecasters. Forecaster A always says:

51%.

Forecaster B distinguishes among:

• 55%;
• 70%;
• 90%;

and those levels are well calibrated. Both can end up correctly classifying a similar number of binary outcomes. But B provides:

more useful information.

Good probabilistic forecasting therefore evaluates both:

• calibration;
• sharpness, or informativeness.

Reviews of probabilistic forecasting emphasize precisely the combination of calibration and sharpness.[4][6]

Superforecasters: good judgment is not only innate talent

The Good Judgment Project followed people who made numerical probability forecasts about future events. Mellers, Tetlock, and colleagues found that the best forecasters maintained high performance across many questions, and that their performance was associated with a combination of:

• cognitive styles;
• specific skills;
• motivation;
• environment and teamwork.[7]

The important lesson is:

calibrated judgment can be learned, at least in part.

One key habit: break a large question into smaller ones. The question:

“Will this organization collapse?”

is too broad. We can break it down:

• What are the financial trends?
• Is it losing members?
• Is there a legal problem?
• Does it have strong succession mechanisms?
• How quickly are revenues changing?

This transforms one large intuitive judgment into several smaller, testable components. It reduces the chance that the whole answer comes from:

one feeling.

Another habit: start with the base rate

Before asking:

“What do I think about this special case?”

ask:

“How often do similar cases usually end this way?”

That is the base rate. Specific information then updates the initial estimate. This is much more stable than beginning with a dramatic story and only later looking for numbers.

A third habit: update gradually. If new information slightly supports a hypothesis, there is no need to jump:

30% → 95%.

A more honest move may be:

30% → 40%.

Bayesian epistemology formalizes the idea that new evidence changes confidence depending on how much more expected the evidence is under one hypothesis than under another.[8] We do not need to calculate every formula. The mental discipline is already useful:

how much should this evidence move me?

Strong evidence and dramatic evidence are not the same thing

A photograph. A recording. Emotional testimony. A shocking document. All can have great psychological impact. But evidentiary weight depends on:

• relevance;
• provenance;
• independence;
• alternative explanations.

So we need to distinguish:

how much the information shocked me

from

how much it should change my probability estimate.

Overconfidence: people are often less precise than they think

Research on probabilistic judgment has long shown that overconfidence is an important problem in human decision-making.[7] But here too we need caution. People are not:

“always overconfident.”

Calibration depends on:

• the task;
• knowledge;
• feedback;
• the way the question is asked.

A more useful rule is:

confidence should be measurable and correctable.

Underconfidence is an error too. If we label every claim:

“maybe”

we avoid being wrong. But we are not very informative. A good researcher does not hide behind:

“nothing is certain.”

They must be able to say:

this is very well documented;

this is probable;

this is open;

this is speculative.

Uncertainty is not an excuse for refusing judgment.

Verbal probability labels are ambiguous. What does:

“probably”

mean? To one person:

60%.

To another:

80%.

Research on uncertainty communication warns that verbal probability expressions are not necessarily interpreted consistently and that numerical, verbal, and graphical formats each have advantages and limitations.[9][10] So THY-REALITY should, when possible, combine:

a verbal label + the reason + where useful, a numerical range.

IPCC as an example of calibrated language

The IPCC developed a structured language for distinguishing among:

• confidence;
• likelihood;
• amount of evidence;
• degree of agreement.[3]

This approach is useful outside climate science too. The important principle is:

do not write only “high confidence.”

Explain:

why.

Confidence and probability are not the same thing. We may estimate:

a 70% probability of an event.

But our confidence in the 70% estimate itself may be:

• high;
• medium;
• low.

For example: a well-validated weather model may produce 70%. Another 70% estimate may be only expert judgment based on very little data. The number is the same. The epistemic quality is not.

False precision. Sometimes it is worse to say:

63.7%

than:

about 60–70%.

If the data do not support decimal precision, numerical exactness creates:

an appearance of knowledge we do not have.

This matters in:

• political forecasting;
• risk assessment;
• historical estimates;
• model projections.

A range may be more honest than a point estimate. If several reasonable models produce:

30–50%,

it may not be useful to calculate:

41.2%

and present it as the true probability. A range may communicate:

model uncertainty

more honestly. The National Academies, in work on decision-making under uncertainty, emphasizes the need to identify, characterize, and transparently communicate different sources of uncertainty.[11]

Under deep uncertainty, one honest probability may not exist. The Stanford review of decision theory also discusses models involving:

• multiple possible probability distributions;
• incomplete preferences;
• confidence-weighted approaches,

precisely because in some real-world problems a single precise probability is not well justified.[2] In such cases we can say:

“I cannot justify one number.”

That is not weakness. It is better calibration.

Expected value is useful, but it does not solve everything. If we have:

• the probability of an outcome;
• the value or harm associated with it;

we can calculate expected value. This lies at the core of standard decision theory under risk.[2][12] But real decisions also involve:

• different tolerance for risk;
• irreversible consequences;
• moral constraints;
• distribution of harms across people.

So mathematical expected value is not always the only relevant criterion.

Small probability × large harm can still matter

If the probability of catastrophe is:

small,

that does not mean:

ignore it.

If the harm is enormous or irreversible, even a low probability can carry decisive weight. Decision theory and the philosophy of risk therefore also examine the limitations of simple expected-utility approaches in catastrophic-risk settings.[2][12]

The precautionary principle is not permission for any fear

The precautionary principle can support action under substantial uncertainty when potential consequences are serious or irreversible. But if used without discipline, it creates another problem:

almost every activity carries some hypothetical risk.

So precaution should still consider:

• plausibility;
• magnitude of possible harm;
• cost of action;
• harms caused by the action itself.

Uncertainty does not eliminate comparison among options.

Decisions are necessary even when the evidence is incomplete. This is an important difference between:

a researcher

and

a decision-maker.

A researcher may say:

we need two more years of data.

A doctor, pilot, crisis manager, or individual may sometimes have to decide:

today.

So:

“the evidence is incomplete”

is not always a reason for:

“do nothing.”

Inaction is also a decision with possible consequences.

Value of information: is it worth waiting for more data?

Decision theory examines the value of information. Additional information is useful if it is sufficiently likely to change the decision and if the value of making a better decision exceeds:

• time;
• cost;
• risk of waiting.[2]

For example: if we must decide in five minutes, a laboratory test available in three days has no operational value for the immediate decision. In another situation, waiting for the test may be the most rational choice.

Information is not always free

Endless checking has a cost. We can:

• spend time;
• miss opportunities;
• become paralyzed.

So rationality is not:

“collect all possible data.”

It is:

“collect enough information given the importance of the decision and the cost of being wrong.”

The opposite danger: premature closure. Medicine, investigation, and everyday reasoning all face the problem of closing a hypothesis too early:

“Now I know what this is.”

Once we accept the first satisfying explanation, we may begin interpreting new information inside it. For important uncertainty it is therefore useful to keep at least:

• the leading hypothesis;
• the strongest alternative.

Two active hypotheses are often better than one

If we are investigating an event, we can write:

H1. the official attribution is correct.

H2

an alternative attribution. Then, for each new piece of evidence, we ask:

which model predicts it better?

This is far more disciplined than:

collecting evidence only for the story we already chose.

Uncertainty is easier to tolerate when it is not an identity. If I say:

“I think H1 is about 70% likely,”

I can revise it after new evidence to:

45%.

If I say:

“I am the kind of person who knows H1 is true,”

the change becomes much more expensive. Probabilistic language therefore helps not only statistically. It helps separate identity from hypothesis psychologically.

Intolerance of uncertainty is a serious research concept, but it should not be abused. Psychology distinguishes among several related constructs:

• intolerance of uncertainty;
• tolerance of ambiguity;
• need for closure;
• uncertainty orientation.

A recent integrative review emphasizes that these concepts are not synonyms and should be distinguished carefully.[13] A systematic review in healthcare found associations between tolerance of uncertainty and various health and professional outcomes, although the literature was methodologically diverse.[14] For our purposes, the key point is simply:

people differ in how uncomfortable they find indeterminacy.

This is not a diagnosis of someone who wants a clear answer.

The need for certainty can change how we use evidence. If uncertainty is very uncomfortable, we may be attracted to an explanation that is:

• clear;
• complete;
• morally simple;

even when the evidence is insufficient. This connects the final foundational article with:

• conspiracy theories;
• high-control systems;
• propaganda;
• expert authority.

All of them can offer the promise:

“We have the answer, so you no longer have to live with ambiguity.”

But endless doubt is not an intellectual virtue either. If we say about every well-documented conclusion:

“What if it is all fabricated?”

we are not more rational. We have simply rejected the possibility that evidence can ever become strong enough. Healthy uncertainty is therefore not:

never conclude anything.

It is:

conclude in proportion to the evidence and preserve the possibility of correction.

Calibration journal

For difficult topics we can use a very simple practice. Before we know the outcome, write down:

• the claim;
• the probability;
• the key reasons;
• what would change the estimate.

Later, check the result. Over time, we get a personal calibration record. That is far more instructive than the memory:

“I am usually right.”

Memory is a poor judge of our old forecasts

After an event, it may feel:

“I knew it.”

This is related to hindsight bias. A record created before the event prevents us from quietly changing:

what we actually predicted.

That is why forecasting tournaments are methodologically interesting:

• the question is clear;
• the deadline is defined;
• the probability is recorded;
• the outcome is later scored.[7]

“I was more or less right” is not good feedback. A forecast must be clear enough for us to know:

what would count as success.

The same applies to a theory. If the definition of success changes after every outcome, we have the same problem as with a self-sealing prophecy. Good uncertainty requires:

a verification rule defined in advance.

Calibrating institutions

We do not have to evaluate only individuals. We can ask:

• How often did an organization's predictions come true?
• Does it publish old predictions?
• Does it measure its forecasting performance?
• Do failed forecasts disappear while successful ones remain in promotional material?

An institution that wants trust should permit:

a track record.

Calibrating the media. A media outlet may write:

“the event is almost certain.”

If it does not happen:

does that ever return in editorial analysis?

Many media systems reward:

• strong forecasts;
• definitive headlines;

less than they reward:

• later measurement of accuracy.

THY-REALITY can work differently here.

Calibrating THY-REALITY. For disputed articles we can introduce:

INITIAL STATUS. the assessment at publication.

NEW EVIDENCE. what appeared later.

STATUS CHANGE. whether the confidence level changed.

REASON

why. That turns the article into:

a living evidentiary structure

rather than:

a monument to the first conclusion.

Do not change your mind because of pressure—change it because of evidence. Flexibility does not mean:

a different opinion every week.

Good updating requires:

• new relevant evidence;
• a change in source quality;
• an improved explanation.

If we change our view only because:

• the mood changed;
• the crowd changed;
• the political climate changed;

that is not epistemic flexibility.

Do not keep your view because of pride either

When new evidence strongly contradicts our earlier position, it is rational to say:

“My earlier assessment was wrong.”

That is not defeat. In a well-designed system:

correction is evidence that the mechanism works.

The degree of revision should follow the weight of new evidence. One anonymous tweet:

little or no movement.

An authentic primary document:

a larger movement.

Several independent documents and admissions by the actors:

more still.

This avoids two extremes:

• every new piece of information overturns the whole theory;
• no information can ever change the theory.

Strong views, loosely held. A useful principle is:

views strong enough to support action, held loosely enough to permit correction.

If we hold a view too weakly:

• we cannot act.

If we hold it as an identity:

• we cannot learn.

The right point lies between the two.

Decision-making under uncertainty also requires values. Two people can accept the same probability:

a 10% chance of serious harm.

One will act. The other will not. Why? Because they value differently:

• risk;
• cost of action;
• freedom;
• safety.

So in a disputed decision we must distinguish:

factual disagreement. from

value disagreement. Otherwise a normative conflict may be falsely presented as a dispute about facts.

Evidence cannot by itself tell us what to value. Data can show:

a measure reduces the risk of X by this much.

They cannot by themselves determine:

how much money, freedom, or other cost is worth paying for that reduction.

That is a political, ethical, or personal question of values. So “follow the science” must not mean:

science itself determines every decision.

Science informs value-based decision-making. It does not replace it.

A transparent decision has three layers. For important topics, THY-REALITY should distinguish:

EVIDENCE. What do the data show?

UNCERTAINTY. How reliable is the estimate?

VALUE JUDGMENT

Which values influence the decision? That is much more honest than hiding all three inside the sentence:

“experts say we must ...”

Uncertainty budget

For every article we can state where uncertainty comes from. For example:

• measurement uncertainty;
• missing data;
• disputed attribution;
• conflicting sources;
• model uncertainty;
• actor intent;
• future developments.

Such an uncertainty budget is more useful than the general sentence:

“there are certain limitations.”

Not every uncertainty matters equally. If we do not know:

the exact hour of an event,

that may not change the main conclusion. If we do not know:

who carried out the act,

that may be the central uncertainty. A good article should therefore say:

which uncertainty could actually change the conclusion.

Robust decisions: what works across several possible worlds? Under deep uncertainty we can ask:

which decision remains acceptable across several realistic scenarios?

Instead of optimizing for:

one most likely future,

we look for:

a robust option.

This is especially useful when model uncertainty is large.

Reversibility: reversible decisions are valuable under uncertainty. If a decision can be:

• tested;
• observed;
• reversed if needed,

we can afford more experimentation. If a decision is:

• irreversible;
• potentially catastrophic,

a higher evidentiary threshold may be justified. This is a practical way to connect:

uncertainty

with

consequences.

Asymmetry of errors. Sometimes two errors have very different consequences.

False positive. We act even though the danger is not real.

False negative

We fail to act even though the danger exists. Which error is worse depends on the problem. So there is no universal rule:

always be more skeptical

or

always be more cautious.

We need a loss function—an understanding of the consequences of both kinds of error.

This is why different fields use different evidentiary thresholds

Criminal proceedings. Medical screening. Scientific publication. Emergency response. Each has a different cost structure for:

• false positives;
• false negatives.

So one universal evidentiary threshold does not make sense for everything. This connects directly to the previous article:

“What Does ‘Proven’ Actually Mean?”

Trust can be graded rather than switched on or off

We do not need to choose only:

I trust / I do not trust.

We can say:

I give this source high weight for technical measurements and less weight for political interpretation.

Or:

I trust this document as authentic evidence of what it says, but not necessarily as proof that every claim inside it is true.

Such a granular model of trust is more rational than:

the entire institution is credible

or

the entire institution lies.

Uncertainty and freedom. A system that promises complete certainty often asks for:

the surrender of part of one's own judgment.

It may be:

• a guru;
• an ideology;
• political propaganda;
• a self-sealing conspiracy theory;
• excessively authoritarian expertise.

Uncertainty is uncomfortable. But it has an important property:

it leaves room for new evidence.

THY-REALITY must not sell certainty

This is the final editorial principle of the ten foundational articles. The project must not say:

“We have the hidden truth that others do not.”

That would reproduce the structure we are analyzing. A better promise is:

“We will show what is well documented, what is probable, what is disputed, what is speculative, and what could change our conclusion.”

It is less dramatic. It is far more worthy of trust.

THY-REALITY UNCERTAINTY PROTOCOL. For every important disputed conclusion:

CLAIM. What exactly are we claiming?

CURRENT STATUS. DOCUMENTED / STRONGLY SUPPORTED / PROBABLE / PLAUSIBLE / SPECULATIVE / UNSUPPORTED / CONTRADICTED

CONFIDENCE BASIS. Why is the assessment at this level?

KEY UNCERTAINTY. Which unknown most affects the conclusion?

ALTERNATIVE. What is the strongest competing explanation?

UPDATE TRIGGER. What evidence would change the status?

DECISION CONSEQUENCE. Must a decision be made despite uncertainty?

REVERSIBILITY. How reversible is the decision?

Calibration discipline. For future articles we add one more rule:

If we use words such as “possibly,” “probably,” “very likely,” or “almost certain,” we should be able to explain what we mean and why we chose that level.

We do not always need to publish a number. But probabilistic language should not be used merely rhetorically.

Conclusion: the goal is not to always be right—the goal is to be correctable

Perfect certainty is rare. But that does not lead to:

relativism.

It is not true that therefore:

all stories are equally good.

We have:

• better and worse evidence;
• better and worse models;
• better and worse calibrated forecasts;
• more and less transparent institutions.

The real goal is not:

never to be wrong.

No:

• person;
• science;
• media outlet;
• institution

can achieve that. The real goal is to create a system in which:

an error can become visible and correctable.

That requires three abilities:

to say “I don't know” when we do not know;

to say “this is very likely” when the evidence is strong;

and

to say “I changed my mind” when new evidence requires it.

That is not weakness. It is the core of intellectual freedom.

Methodological note

This article does not claim that:

• all confidence levels must be expressed numerically;
• expected utility solves every decision problem;
• the Brier score measures every dimension of judgment quality;
• superforecasting means the future is precisely predictable;
• intolerance of uncertainty explains every desire for a clear answer;
• the precautionary principle always requires the stricter action;
• uncertainty means all options are equally likely.

Probability and decision-theory models are used as tools for more transparent judgment, not as replacements for moral, political, or personal values.