Context

There may be no sentence more repeated, and less examined, in modern technical life than this one. It is the epigraph of data science, the first slide of machine-learning lectures, the resigned shrug of every analyst who has watched a beautiful model fail in production and a crude one succeed. It is usually delivered as a joke and usually received as one, which is a shame, because it is a complete epistemology in eleven words. Its author was George E. P. Box, a British statistician who is one of the few people of the twentieth century whose name is attached both to the tools everyone uses and to the warning about using them. The line is often treated as an excuse. Read properly, it is a standard.

The attribution is right; the citation almost never is. Almost everyone credits the quote to Box’s 1976 paper “Science and Statistics,” published in the Journal of the American Statistical Association (vol. 71, no. 356, pp. 791–799). What that paper actually contains, on page 792, is only the first half: “Since all models are wrong the scientist cannot obtain a ‘correct’ one by excessive elaboration,” and, two paragraphs later, “Since all models are wrong the scientist must be alert to what is importantly wrong.” The reassuring second clause was not there yet. The complete sentence, “All models are wrong but some are useful,” appears as a section heading in Box’s 1979 chapter “Robustness in the Strategy of Scientific Model Building,” in the volume Robustness in Statistics edited by Robert L. Launer and Graham N. Wilkinson, based on a 1978 statistics workshop. It surfaces again, in the form most people quote, on page 424 of Box and Norman Draper’s Empirical Model-Building and Response Surfaces (1987): “Essentially, all models are wrong, but some are useful.”

Box is worth knowing because the aphorism is not a philosopher’s quip. He was born in Gravesend, Kent, on 18 October 1919, left school at sixteen, and was working as an assistant at a sewage treatment plant when war broke out. Assigned to chemical-defence work, he found that the results of his experiments on poison gas were too variable for his chemistry to explain, so he taught himself statistics from the available texts and never looked back. He took a BSc in mathematics and statistics at University College London in 1947 and a doctorate from the University of London under Egon Pearson and H. O. Hartley, then spent years as an industrial statistician at Imperial Chemical Industries before moving to Princeton University to direct its Statistical Techniques Research Group. In 1960 he founded the Department of Statistics at the University of Wisconsin–Madison; he was elected a Fellow of the Royal Society in 1985 and died in Madison on 28 March 2013, aged 93. He spent his whole life building and repairing models. When he says all of them are wrong, he is not being modest. He is being precise.

Meaning

The word “wrong” is doing careful work, and misreading it is the most common way to misread the quote. Box does not mean that models are fraudulent, or that their authors are dishonest, or even that they are usually inaccurate. He means something structural. A model is a deliberately simplified stand-in for something too complicated to hold in the mind at once, and to be a stand-in it must leave things out. Omission is not a defect that better data or more parameters can fix; it is the price of abstraction itself. A perfect model of the weather would be the weather. The moment you commit to a representation, you have chosen what to ignore, and everything you ignored is a way the model can be wrong. On this view, error is not a bug in models. It is their defining feature.

The pivot of the sentence — the small word “but” — then moves the test from truth to use. Since no model can be correct, the sensible question is not “is it true?” but “is it useful, and for what?” Box made the test explicit elsewhere in the same work: “Remember that all models are wrong; the practical question is how wrong do they have to be to not be useful” (Empirical Model-Building and Response Surfaces, 1987, p. 74). That is a boundary condition, not a slogan. Usefulness is purpose-relative; the same model can be a triumph for one decision and a catastrophe for another. The 1976 paper supplies the practical corollary: because you cannot buy correctness with complexity, you should prefer the most economical model that does the job. The aphorism is therefore also an argument for parsimony, not just for humility.

The two halves are in tension, and the tension is the teaching. “All models are wrong” is a guard against idolatry: never confuse the map for the territory, the metric for the goal, the forecast for the future. “But some are useful” is a guard against the opposite error: if every model is wrong, the cynic concludes that none is worth building, and abandons measurement altogether. Box’s sentence holds both truths at once, which is why it cannot be paraphrased into either half alone. There is a final, self-referential turn: the aphorism is itself a model of modelling, and by its own terms it too is a simplification. It is a very useful one. It is not the last word, and Box would have been the first to say so.

Relevance

The quote has become the intellectual weather of the artificial-intelligence era, and for good reason. Large language models are the most vivid illustration of Box’s point ever built: they are fluent, confident and useful on an extraordinary range of tasks, and they are also wrong in ways that fluency actively hides. A grammar error would be easy to catch; a plausible fabrication delivered in perfect syntax is not. Box’s framework predicts this exactly. The problem is not that the model occasionally errs and could be patched into truth; it is that error is intrinsic to a system whose entire job is approximation. The mature response is therefore not to ask whether the model is trustworthy in the abstract, but to ask what it is being trusted to do, how often it fails at that specific task, and how you would find out. Both the hype that treats models as oracles and the dismissal that treats them as toys miss the same distinction.

The same discipline applies far beyond software. Forecasts, credit scores, risk models, school rankings, benchmarks and dashboards are all models, and all of them are wrong in ways that are useful until the day they are not. The two failure modes Box’s sentence prevents are opposites and equally dangerous. One is model worship: treating the output of a simplified system as a description of reality, and then optimising blindly against it until the measure detaches from the thing it measured. The other is model nihilism: rejecting expertise, measurement and forecasting because none of them is perfect. The line also explains a specific modern hazard — a model validated in one domain is often deployed in another where its omissions are suddenly the whole story, so that its usefulness expires without warning. Utility is local. Wrongness is universal.

Application

The practical way to use the quote is to attach three questions to every model you rely on, whether it is a machine-learning system, an economic forecast or your own back-of-the-envelope estimate. First, what is this model for? Second, what is it assuming away? Third, what would tell me it has stopped being useful? The last question is the one Box added, and it is the one people skip: define the failure signal in advance, so that you notice the model has outlived its purpose rather than discovering it in the post-mortem. For high-stakes decisions, triangulate with an independent method, keep a human judgement in the loop, and resist asking a tool for a certainty it was never designed to provide. You are not looking for a true model. You are looking for one whose known errors are acceptable for the purpose at hand, and a way to know when that stops being true.

The second application is inward. You run models of your own all day: forecasts about your career, narratives about your relationships, plans, budgets, beliefs about what you are capable of. “All models are wrong” grants permission to act without certainty, because certainty was never on offer. “Some are useful” imposes the harder obligation: keep the ones that earn their keep, retire the ones that do not, and judge them by their results rather than their elegance. Most of us do the reverse, defending our models because they are ours. The instruction is to hold them firmly enough to use and loosely enough to revise — to ask, of a belief or a plan, how wrong it would have to be before it stopped being useful to keep. The danger is not being wrong. It is being wrong in a way your model was built to conceal.

Create your own quote graphic with this quote

This is a line that rewards a clean setting — short enough to read at a glance, dense enough to keep thinking about. When you pair it with simple, confident typography, you get something worth pinning above a workstation. Use our free online quote maker to turn this quote into a shareable image in seconds — no signup needed. For advanced features like video quotes and 100+ templates, try the Quotes Creator Android app.