Roll a fair six-sided die and I can tell you everything about the gamble before it lands: each face has probability one-sixth, the average roll is 3.5, and if you pay me a dollar to win six when a six comes up, I can price that bet to the penny. Now I ask you a different question: what’s the probability that a technology no one has invented yet will dominate your industry in twenty years? You can’t answer — and neither can I, and neither can the smartest forecaster alive. Not because we’re lazy, but because the number does not exist to be looked up.
Those are two fundamentally different animals wearing the same coat labelled “the future is uncertain.” Telling them apart is the first and most important skill in this entire course, because the tools that master one are actively dangerous on the other. The economist who drew the line got there in 1921.
Knight’s distinction: measurable vs. unmeasurable
In his book Risk, Uncertainty and Profit, economist Frank Knight made a distinction so useful it now carries his name. He split “we don’t know what will happen” into two categories:
- Risk is measurable uncertainty — a gamble whose probabilities are known or knowable. You can put a number on each outcome. A fair die, a roulette wheel, a coin flip, a life-insurance table built from millions of deaths, the failure rate of a mass-produced part. Here the world hands you the odds, or lets you estimate them from abundant, stable data. This is the home turf of expected value.
- Uncertainty (Knightian uncertainty, and at its extreme, deep uncertainty) is unmeasurable — a gamble whose probabilities are unknown, contested, or unknowable. Which startup wins a brand-new market, whether a war breaks out, how a never-before-seen virus spreads, what the economy does in a decade. There’s no stable data to count, no symmetric mechanism to reason from, no reference class you fully trust. The number you’d need simply isn’t available.
The two in one line
Risk: you know the odds (a die — each face is 1-in-6). Uncertainty: you don’t, and often can’t (which technology wins the decade — no probability to look up). The mistake that wrecks decisions is treating the second as if it were the first: inventing a number and then optimising it as though it were real.
The tell is where the probability comes from. For risk, it’s either handed to you by a symmetric mechanism (a die has six equal faces by construction) or harvested from a mountain of stable, repeatable data (insurers have seen millions of lives). For deep uncertainty, there’s no such source — you’re facing a one-off event with no reliable reference class, so any probability you write down is really a feeling dressed in a decimal point.
Which of these is a matter of RISK (knowable odds), not deep uncertainty?
A spectrum, not a switch
It’s tempting to sort every decision into two clean bins, but reality is a dial. At one end sits pure risk (the casino, where odds are exact). Slide along and you hit estimable risk, where you don’t know the odds exactly but have enough stable data to estimate them well (the failure rate of a bridge design, the churn rate of a mature subscription business). Keep sliding into deep uncertainty, where data is thin, the situation is novel, and estimates are little more than guesses (a new market, a new drug’s societal effects). At the far end lies genuine ignorance or “unknown unknowns” — outcomes you haven’t even imagined, so you can’t assign them a probability because you don’t know they’re on the menu.
The practical point isn’t to name the exact spot on the dial. It’s to notice, for any given decision, how much of your probability is real and how much is invented — because that ratio tells you how far to trust an expected-value calculation and how hard to lean on the robustness tools this course is about.
| Risk | Deep uncertainty | |
|---|---|---|
| Probabilities | Known or well-estimated | Unknown, contested, or unknowable |
| Source of the odds | Symmetric mechanism or abundant stable data | No reliable reference class; one-off events |
| Example | A die, roulette, a mortality table | Which tech wins the decade; a novel pandemic |
| Right tool | Expected value, optimise the average | Robustness, margin of safety, survival |
| Failure mode | Miscalculating a knowable number | Inventing a number and trusting it |
Sort each decision by the kind of unknown it really is.
Place each item in the right group.
- An insurer pricing a 60-year-old's one-year mortality from millions of records
- How a technology nobody has built yet reshapes an industry over 20 years
- The long-run fraction of red cards drawn from a well-shuffled deck
- Whether a start-up in a brand-new category becomes the dominant platform
- The probability two dice sum to seven
- Whether an unprecedented financial crisis strikes in the next decade
Why the confusion is so dangerous
If deep uncertainty just meant “we’re less sure,” the fix would be to widen your error bars and carry on. The real danger is subtler and worse: our tools for risk don’t refuse to run on uncertainty. Plug a made-up probability into an expected-value model and it returns a crisp, confident, decimal-pointed answer — one that looks exactly as authoritative as a correctly-priced insurance policy. The rigour is entirely cosmetic. Garbage odds in, beautifully-formatted garbage out. This is false precision, and it’s the villain of the next lesson.
Worse, treating uncertainty as risk quietly encourages fragile choices. If you believe your probabilities are solid, you’ll happily optimise hard against them — concentrate your bets, strip out the “wasteful” safety margin, lever up to squeeze the expected value. Every one of those moves is exactly right under real risk and potentially fatal under deep uncertainty, where the future you optimised against may simply never show up.
Because not deciding is itself a decision — usually the fragile one. Refusing to choose until the fog clears means defaulting to the status quo, which is just an un-chosen bet on “things stay the same,” often the least robust option of all. Deep uncertainty is the normal condition of every important choice — careers, investments, strategy, health, relationships. If you waited for knowable odds you’d never act at all. The skill this course teaches is precisely how to decide well anyway: not by manufacturing fake probabilities, but by choosing actions that survive whichever future actually arrives. You don’t beat the fog by waiting for it to lift. You beat it by packing for every kind of weather.
Fill in Knight's distinction.
Pick the right option for each blank, then check.
Under , the probabilities are , so you can compute an expected value and optimise it. Under deep , the probabilities are unknown or unknowable, so plugging a made-up number into the same model produces — a confident answer built on numbers you invented.
When to reach for this distinction
Run this check at the start of any consequential decision, before you open a spreadsheet: where do my probabilities actually come from? If they flow from a symmetric mechanism or a deep well of stable data, you’re in risk — compute away, expected value is your friend. If you’re quietly making them up because the situation is novel and one-off, you’re in deep uncertainty — and the honest move is to stop optimising the fake numbers and switch to the robustness toolkit: cap your downside, keep options open, and choose what survives a wide spread of futures. Naming which world you’re in is half the battle, because the two demand opposite instincts, and the costliest mistakes come from using the risk instinct in an uncertain world.
Next up: When Expected Value Lies — we take the workhorse of the whole latticework, the expected-value average, and show exactly where and why it betrays you under deep uncertainty. You’ll meet ergodicity, ruin, and an interactive many-worlds explorer that lets you watch the “optimal” strategy blow itself up.