Knightian Uncertainty

When rolling dice, you know exactly what you’re dealing with. The six comes up in one out of six cases, and that can be calculated. Economist Frank Knight calls such situations “risk”: the outcome is open, but the probabilities are known. The situation is entirely different when it comes to the question of how many people will have lost their jobs to AI in ten years. Here, no one knows the probabilities, because there is no track record from which to derive them. It is precisely this second kind of not-knowing that has been called Knightian uncertainty since 1921.

Why models hit a limit here

Almost all of financial mathematics presupposes known probabilities. An insurance company can calculate car accidents because it knows millions of accidents from the past. From this data comes a premium, and on average the math works out. But as soon as an event is unprecedented, this method breaks down. There is no past from which to extrapolate the future.

The mistake usually doesn’t happen during the calculation, but before it. Uncertainty is treated as if it were risk, and the result is a number with many decimal places. The number looks credible, even though the assumptions behind it are pure invention. In the 2008 financial crisis, this was a central problem: risk models for mortgage loans relied on decades in which house prices had never fallen across the board. The case that actually occurred simply did not appear in the data.

Knight drew an economic conclusion from this. Genuine profit is only made where this kind of uncertainty prevails. Calculable risk can be hedged by anyone, which is why it yields hardly any return. Whoever makes the right decision under true uncertainty, however, gets paid for it — this is Knight’s explanation for why entrepreneurs exist.

Risk, uncertainty, and the difference in everyday life

The dividing line runs along the question of whether the probabilities are known. In roulette, they are known and unchanging. With the weather, they are not known exactly, but they can be estimated fairly well from a very large number of past weather patterns. When it comes to whether a particular AI company will still exist in five years, there is no comparable body of data. You can name a number, but it’s an opinion, not a measurement.

A second difference concerns the list of possible outcomes. With a die, you know all six outcomes. With a new technology, you don’t even know the full list. Before 2020, hardly anyone would have put “models write functioning program code” on such a list. If even the possible outcomes are unknown, a probability calculation is pointless from the outset.

In practice, this is why uncertainty is handled differently from risk. Instead of calculating an expected value, one works with scenarios and buffers. The question is not “how likely is the worst case” but “will I survive it if it happens”. This attitude is called robustness: the decision should remain viable even if the assumptions turn out to be wrong.

AI valuations and regulation as illustrative material

In reporting on AI, the term comes up regularly in discussions of company valuations. A start-up with no significant revenue is valued at many billions. Such figures are not based on a calculation but on guesses about a market that does not yet exist. Critics speak of bubble formation, proponents of a bet on a new era. Both sides have the same set of data — which is why no calculation can settle the dispute.

The topic is also present in regulation. Laws such as the EU AI Act must set rules for systems whose future capabilities no one knows. That is why they contain reporting obligations, reviews, and revision clauses instead of fixed thresholds for everything. This is exactly the robustness strategy: mechanisms are built in that can react, rather than committing to a single forecast.

A common misconception is that Knightian uncertainty means you can’t know anything at all. That’s not true. You can estimate boundaries, define early warning signs, and keep decisions reversible. The term merely warns against trusting numbers that feign a certainty that doesn’t actually exist.