Risk and Uncertainty Are Different

Imagine a fair six-sided die before it is thrown.

You do not know which face will appear. Yet under the accepted fair-die model, you can specify the possible results and assign each one a probability. The outcome is unknown, but the structure of the situation is represented.

Now imagine launching a product into a market that does not yet exist in a familiar form. You may estimate demand, but you may not know which competitors will enter, how customers will understand the product or which developments will change the market itself.

Both situations involve an unknown future. They do not involve the same kind of knowledge.

The first is primarily a situation of risk. The second contains deeper uncertainty.

An unknown result can still be structured

Risk does not mean knowing what will happen.

With the die, the next result remains unknown. What is represented—under the stipulated fair-die model—is the set of possible outcomes and the probability assigned to each one.

This makes probabilistic calculation possible. It does not make the result certain or controllable.

Even this simple classification depends on assumptions. The die must be sufficiently fair. The throw must not be manipulated. The possible outcomes must be represented correctly. If those assumptions become doubtful, the decision contains model uncertainty in addition to the ordinary risk of which face will appear.

An exact probability derived from an accepted model is not necessarily an exact probability known to describe the real physical process.

The distinction is therefore not between an unknown future and a known future. It is between different structures of incomplete knowledge.

What risk means

In its narrow decision-theory sense, risk concerns specified outcomes whose likelihoods can be represented or defensibly estimated.

The probability basis may come from:

• a known physical structure;
• repeated observations from a sufficiently stable process;
• a statistical model with relevant empirical support;
• informed probabilistic judgment.

“Defensible” is relative to the available evidence, the model, the purpose of the estimate and the stability of the represented process. Risk requires enough structure to represent variation. It does not require perfect knowledge.

Frank Knight’s influential account distinguished measurable risk from uncertainty that could not be assigned an objective probability. His central insight was that an unknown result inside a classifiable process differs from a situation whose probability structure cannot be measured in the same way (Knight, 1921).

Modern decision theory does not use Knight’s terminology uniformly. Later frameworks allow beliefs to be represented through subjective probabilities even when objective frequencies are unavailable. Contemporary risk disciplines also sometimes use risk more broadly to include consequences, uncertainty and the strength of the supporting knowledge.

Knight’s distinction remains useful, but it is not the only modern taxonomy.

Writing a number beside an outcome does not, by itself, make the situation one of calculable risk.

Probabilities make different claims

A specified probability describes an accepted formal or physical process. Under the fair-die model, each face receives a probability of one in six.

A frequency estimate is inferred from repeated observations. Its usefulness depends on sample quality and on whether the observed process remains relevant to the future.

A model-derived probability is produced from data and structural assumptions. Its credibility depends on the model’s inputs, calibration and relevance to the situation.

A subjective probability represents a degree of belief. It can incorporate information that is difficult to reduce to observed frequencies, but it should not be confused with an objectively established chance.

Subjective does not necessarily mean arbitrary. A judgment can be informed, internally coherent and responsive to evidence. In Savage’s subjective expected-utility framework, preferences satisfying specified consistency conditions can be represented through utilities and subjective probabilities. This is a formal representation under assumptions, not proof that every expressed confidence level is coherent or accurate.

These are sources or interpretations of probability, not mutually exclusive containers. A model-derived probability may use observed frequencies. A subjective judgment may incorporate model outputs.

All can support decisions. They do not carry equal evidential authority.

What uncertainty means

Uncertainty is broader than not knowing which specified outcome will occur.

A decision may remain uncertain because:

• relevant outcomes have not been identified;
• the probabilities of known outcomes are unclear;
• important variables are missing;
• causal relationships are poorly understood;
• several models remain plausible;
• the environment may change;
• the situation can develop in ways absent from the representation.

A person can possess substantial information and still face uncertainty. The term does not require total ignorance, nor does it prevent estimation.

Partial evidence may constrain what is plausible without supporting one complete probability distribution. Knowledge can improve in one dimension while remaining limited in another.

The central question is not simply whether a number exists. It is what the number represents, which assumptions support it and what remains outside it.

Ambiguity is one kind of uncertainty

Ambiguity commonly refers to decisions in which the outcomes are known but their probabilities are missing, imprecise or unreliable.

In an Ellsberg-type choice, one container holds known proportions of differently coloured balls. Another contains the same possible colours, but their proportions are unspecified. People frequently prefer to bet on the container with specified probabilities, producing a pattern difficult to reconcile with standard subjective expected utility (Ellsberg, 1961).

This is commonly called ambiguity aversion.

But ambiguity aversion is not universal. Experimental evidence also finds ambiguity neutrality and systematic ambiguity seeking. Responses vary across uncertainty sources and contain individual-specific components (Li et al., 2018).

They can also vary with the degree of ambiguity, including acceptance of lower ambiguity alongside avoidance of higher ambiguity (Klingebiel and Zhu, 2023).

Ambiguity is narrower than uncertainty. The possible colours in the container task are known. In more open situations, the difficulty may be that relevant outcomes, variables or causal relationships have not been identified.

Unknown probabilities are one limit on knowledge. They are not the only one.

A model can quantify one layer and miss another

A model can represent risk precisely while remaining uncertain about its own adequacy.

Several layers can be distinguished.

Outcome variability concerns which specified outcome will occur.

Parameter uncertainty concerns the correct numerical values inside an accepted model. A demand model may be structurally accepted while the expected response rate remains uncertain.

Model uncertainty concerns which model adequately represents the process. Several candidate models may fit earlier observations while predicting different futures—and the candidate set itself may be incomplete.

Structural uncertainty concerns whether relevant variables, states or relationships are missing.

Beyond these lies the possibility of outcomes that have not been represented at all.

Decision-modelling research distinguishes uncertainty about parameter values from uncertainty about structural assumptions. Assigning probability distributions to parameters does not automatically account for whether the model contains the right relationships (Jackson et al., 2011).

Likewise, a precise point estimate is not a complete account of uncertainty surrounding a modelled decision. Quantitative outputs remain conditional on structural and methodological choices (Briggs et al., 2012).

These layers are an explanatory scaffold, not a taxonomy used identically across disciplines. Their function is to show that a calculation can be exact relative to its assumptions while the assumptions remain uncertain.

Precision is not the same as knowledge

Suppose a forecast assigns an event a probability of 31.7%.

The calculation may genuinely produce that value. Its evidential strength still depends on:

• the quality and relevance of the data;
• uncertainty in the parameters;
• the suitability of the model;
• the stability of the environment;
• whether important possibilities were omitted.

A highly specific estimate is not automatically false precision. False precision arises when the specificity of the reported number exceeds what the evidence and modelling assumptions can support.

This is the difference between computational precision and evidential precision.

Quantification remains valuable. It can expose assumptions, permit comparison, support calibration and show how conclusions change when inputs change. The error is treating numerical specificity as proof that the underlying knowledge is equally specific.

Sometimes evidence supports a probability range or set rather than one point estimate. Imprecise-probability approaches include intervals, lower and upper probabilities and sets of distributions. They represent the idea that evidence may constrain probability without identifying one uniquely defensible value (Cozman, 2017).

A range is not automatically more honest. Its boundaries may also be poorly supported, and it does not resolve uncertainty about missing outcomes or model structure.

Numbers can organise uncertainty. They do not abolish it.

Past frequencies depend on a stable-enough world

Consider the launch of an unfamiliar product.

Historical data may describe related products, customer behaviour or previous market growth. These observations can support estimates, but applying them assumes enough continuity between the earlier environment and the emerging one.

That assumption may weaken when the process generating outcomes changes.

A competitor can enter. Technology can alter production or distribution. Customers can interpret the product differently from anything in the historical data. Regulation, pricing or social meaning can change the market structure.

Past evidence does not become useless whenever conditions change. Earlier observations may still constrain plausible expectations. Their relevance is simply no longer automatic.

A historical frequency describes what occurred under earlier conditions. Transporting it into a new environment introduces uncertainty about whether those conditions remain comparable.

The product launch can therefore contain calculable components without becoming entirely calculable risk.

Real decisions contain both

Suppose an outdoor event depends on tomorrow’s weather.

A forecast may assign a 30% probability of rain. That number represents modelled weather variation and is conditional on a forecasting system, issuance time and geographic representation. It can inform the decision without revealing what will happen.

Additional uncertainty may remain.

The forecast may be less well calibrated or less locally informative for the precise location. Conditions may change after it is issued. Different models may disagree. The event’s vulnerability may depend on wind, timing or ground conditions not captured by the headline probability.

The decision contains represented risk alongside uncertainty about model adequacy, local conditions and relevant consequences.

Calling the situation risky is not necessarily wrong. Calling it uncertain is not necessarily wrong either. Each term identifies a different layer.

The useful question is:

Which part of the situation has been represented, and which part remains inadequately known?

Mixed cases do not erase the distinction. They show why it is needed.

Risk is not danger or control

In ordinary language, risk often means danger. The concepts overlap, but they are not identical.

Danger concerns the possibility of harm. Risk concerns possible outcomes and their represented likelihoods.

An outcome can be extremely dangerous but have a low estimated probability. Another can be highly probable but cause little harm. Saying that something is “high risk” remains unclear unless it identifies whether the concern is:

• probability;
• consequence;
• expected loss;
• strength of supporting knowledge;
• some combination of these.

Risk is also not controllability.

A process can be statistically well described while remaining beyond the decision-maker’s control. Weather can be forecast without being controlled. Conversely, a person may influence an unfamiliar situation without possessing reliable probabilities for its outcomes.

Prediction and control can interact, but neither guarantees the other.

What the distinction clarifies

Risk and uncertainty are not two interchangeable names for any unknown future.

Risk concerns possible variation inside a represented structure. Outcomes and their likelihoods can be specified or estimated on some defensible basis.

Uncertainty also concerns the limits of that structure: missing probabilities, uncertain parameters, competing models, incomplete causal relationships, unstable conditions and possibilities that have not been represented.

Real decisions often contain several of these layers at once.

Separating them clarifies:

• what is specified;
• what can be calculated;
• what must be estimated;
• what depends on assumptions;
• what remains unknown.

It does not determine what action should be taken.

Risk describes possible variation inside a represented structure. Uncertainty also includes what that structure cannot yet represent reliably.