CONCEPT
The Survivor's Error
Abraham Wald’s structural insight that a sample filtered by a process correlated with the outcome you care about cannot be analyzed its way to the truth—the error that every training dataset in machine learning inherits and that scale reliably compounds rather than dissolves.
The most consequential invisible failure mode in deployed machine learning has a name that predates the field by eighty years.
Abraham Wald identified it in 1943 while estimating aircraft vulnerability from damage on returning bombers: the planes that came back were not a random sample of all planes, but a sample conditioned on survival, and survival was correlated with exactly the property he was trying to measure. The popular version of the story—look where the holes aren't—keeps the punchline and discards the proof. The proof is what matters: Wald showed that you cannot analyze your way out of the Survivor's Error by examining the sample more carefully, because the error is not in the analysis but in the structure of the sample. The absent planes are absent precisely because of what you want to know about them. The only escape is to import a model of the selection process from outside the