All Models Are Wrong
Why imperfection is not a failure of modelling, but the condition that makes models usable.
“All models are wrong, but some are useful.”
— George E. P. Box
A perfect model of the universe would need to be the universe.
Every practical model compresses. Newtonian mechanics ignores relativistic effects at ordinary speeds. A weather model divides a continuous atmosphere into finite cells. A neural network learns statistical regularities rather than carrying a miniature copy of reality inside it.
So the important question is not simply:
Is the model true?
It is:
Is it accurate enough, in the ways that matter, for this purpose?
A subway map distorts distance but preserves connectivity. A linear regression may miss intricate curvature yet reveal a dominant tendency. Word2Vec does not understand a word as a person does, but its geometry can still expose useful semantic relationships.
“Wrong” therefore does not mean worthless. It means conditional. Every model comes with a domain, a scale and a set of assumptions. Wisdom lies in knowing where its usefulness ends.
Source note: The aphorism is associated with statistician George E. P. Box and appears in his writing on model-building and scientific method.