Risk and Uncertainty

The idea in plain words

A casino knows the odds on every bet. An insurance company knows roughly how many houses will burn down each year. These are risks: the possible outcomes are known and their frequencies can be estimated. Now consider a company deciding in 1995 whether to bet on the World Wide Web, or a team in 2023 deciding how generative AI will change their product. Nobody could list the possible outcomes, let alone assign them probabilities. That is uncertainty.

The distinction matters because the tools are different. Risk can be priced, diversified and insured. Uncertainty cannot, or not in the same way. Many decision failures come from using risk tools, such as expected value calculations, detailed forecasts and sensitivity tables, on problems of uncertainty, where they produce numbers that look precise and mean little.

Where it comes from

The economist Frank Knight drew the line in 1921. Risk, he argued, refers to situations where outcomes follow a known or estimable distribution, either because we can calculate it, as with dice, or because we have enough similar past cases, as with fires. Uncertainty refers to situations that are unique, where no such grouping is possible. Knight’s striking conclusion was that profit itself comes from uncertainty. If all futures were measurable risks, competition would price them away. Entrepreneurs earn profit by making judgements where no calculation is possible.

John Maynard Keynes reached a similar view from philosophy. His Treatise on Probability (1921) argued that many probabilities cannot be expressed as numbers, and some cannot even be compared. In 1937, defending his General Theory, he listed examples of genuine uncertainty: the prospect of a European war, the price of copper twenty years hence, the obsolescence of a new invention. About such matters, he wrote, “We simply do not know.” Since people must act anyway, they fall back on conventions: assuming the present will continue, following the opinion of others, relying on the average judgement of the market.

In 1961 Daniel Ellsberg showed that people feel the distinction even when theory says they should not. Offered a bet on drawing a red ball from an urn with a known half-red mix, or from an urn with an unknown mix, most people prefer the known urn, whichever colour they are betting on. This is ambiguity aversion: we dislike not knowing the odds more than we dislike bad odds.

What it does to the mind

Uncertainty is uncomfortable, and the mind relieves the discomfort by converting it into risk. A forecast with a number attached feels more manageable than an admission of ignorance. Organisations formalise this conversion: business cases demand projected revenues, five-year plans demand market sizes. The numbers then acquire an authority their foundations do not support.

Ambiguity aversion also biases choices towards the familiar. When two options have similar expected value, people pick the one whose risks they can describe, which often means the incumbent technology, the existing market or the known competitor. The genuinely new option is penalised not for being worse but for being unknown.

Where it shows up

The compute essay describes decades in which the industry treated steady performance gains as a measurable trend, a risk to plan around, until physical limits turned it into an uncertainty. The word processors essay ends in the uncertainty created by AI, where no one knows what writing will be for. Gates’s Internet Tidal Wave memo is a response to uncertainty: instead of a forecast, it commits the company to a direction that would pay off across many possible futures.

Iridium is the opposite. Its business case treated a decade-long future of mobile phone coverage as something that could be forecast, and built a fixed, irreversible system around the forecast.

What it asks of someone deciding

Ask which kind of problem you face. If there is a reference class of similar past cases, use it. If there is not, be suspicious of any number that claims to describe the future precisely.

Under uncertainty, prefer robustness to optimisation. Choose actions that are acceptable across many futures, keep options open where it is cheap, and make commitments in stages that let you learn.

Be honest in documents. “We do not know, and here is how we will find out” is a better basis for a decision than a forecast whose confidence nobody believes.