What Is the Problem of Induction? (Taleb's Black Swan Foundation)

The problem of induction is the philosophical observation, formalized by David Hume in the 18th century, that no finite set of confirming observations can logically prove a universal rule.

You observe 1,000 white swans. Can you conclude all swans are white? Logically: no. The 1,001st swan might be black. Any finite sample, no matter how large, leaves open the possibility that the next observation will disconfirm the rule.

One black swan disproves the rule "all swans are white" with certainty. One million white swans do not prove it.

The Asymmetry That Matters

Hume established that induction is logically invalid. Karl Popper drew the practical consequence: scientific knowledge grows through falsification, not confirmation. A valid scientific claim is one that can be falsified — one that specifies what observation would make it wrong. Theories that survive serious attempts at falsification are corroborated, not proven.

Nassim Taleb takes this asymmetry into risk management. If no number of confirming observations proves safety, then long track records of safe periods do not prove safety of a strategy. They prove it has survived the observations so far — which says nothing about whether it can survive the next observation.

The Turkey Illustration

A turkey is fed every day for 1,000 days by the farmer. Each day adds to the turkey's evidence that the farmer's intentions are benevolent. The evidence is strongest on day 999. On day 1,000 (Thanksgiving), the turkey's confidence-building data set is dramatically disconfirmed.

The turkey's error is induction: concluding from 1,000 confirming days that the pattern will hold. The farmer wasn't changing — the observation was selecting the safe period before the reversal.

The Risk Management Application

Strategies that look safe because they've been safe are not safe. The absence of a disconfirming event in the sample does not establish the absence of the event from the distribution.

For risk management, this means: ask not just whether the bad event has occurred in the data, but whether there's any structural reason it couldn't. If there's no structural reason — if it's absent from the sample only because the sample ended before it arrived — the sample's track record provides no protection against it.

For the full framework, read Living With Randomness.