All Models Are Wrong, but Some Are Useful

In 1976 the statistician George Box wrote a line that has become the perfect companion to Korzybski’s: “All models are wrong, but some are useful.” Box was talking about statistical models, but the sentence escaped the lab and now applies to every map you think with.

It’s a deceptively dangerous sentence, because there are two ways to read it and beginners almost always grab the wrong one.

The wrong half and the right half

The wrong takeaway: “All models are wrong, so models are useless — just look at reality directly.” This sounds tough-minded and is actually impossible. You can’t look at reality directly; reality is the territory, and it’s far too big and detailed to hold in a human head. The moment you try to think about anything at all, you reach for a map. Refusing all models doesn’t connect you to raw truth; it just leaves you using maps you refuse to admit you’re using.

The right takeaway: Since every model is wrong somewhere, your job is not to find the one perfect model. It’s to (1) choose a model that’s useful for the question in front of you, and (2) stay awake to where that model is wrong. “Wrong but useful” isn’t a verdict against models — it’s the correct way to hold every single one. A weather forecast is wrong (it won’t nail the exact temperature) and enormously useful (bring the umbrella). Both halves are true at once, and that’s fine.

The same pattern, in four very different maps

The reason this idea is foundational is that everything you reason with turns out to be a map — not just literal maps and weather forecasts. Watch the identical pattern (useful structure preserved, crucial detail dropped) appear in four domains that look nothing alike.

1. A financial model is not the business

A spreadsheet model of a company captures revenue, costs, and growth in tidy cells. It’s genuinely useful — you can test “what if we raise prices 10%?” without risking real money. But it silently drops the founder’s burnout, the key engineer about to quit, the lawsuit nobody’s mentioned, and the customer whose loyalty is purely habit. The model says “the business”; it means “the parts of the business that fit in cells.” Mistake one for the other and you’ll be blindsided by the thing that was never in the spreadsheet.

2. A stereotype is not the person

A stereotype is a brutally compressed map of a group of people. Like all maps it’s a generalisation that’s sometimes structurally useful (you do, sensibly, expect a stranger in a pilot’s uniform to know how to fly). But applied to an individual it’s wrong in the way that matters most: a person is a territory, full of detail the category never recorded. The error isn’t having categories — your brain can’t function without them — it’s treating the category as the person. The map is not the territory; the group label is not the human.

3. GDP is not wellbeing

Gross Domestic Product is a map of a country’s economy: the money value of what it produces. Hugely useful, and it does mirror real structure — richer countries genuinely tend to have better health and schools. But GDP cheerfully counts a traffic jam’s extra fuel and a hurricane’s rebuilding as “growth,” while ignoring unpaid care work, clean air, and whether anyone is actually happy. Govern as if GDP were national wellbeing and you’ll optimise the map while the territory — people’s actual lives — drifts somewhere the number never looked.

4. The org chart is not the organisation

A company’s org chart maps who reports to whom. Useful for “whose budget is this?” — and almost a lie about how work really happens. The real organisation runs on a hidden territory of friendships, favours, the one veteran everyone secretly asks, and the back-channel where decisions actually get made. New managers who trust the org chart as the territory keep wondering why pulling the official levers does nothing.

Why “wrong” doesn’t mean “useless”

Put the four cases together and a rule pops out: a map being wrong tells you nothing about whether it’s useful. Usefulness is a separate question — does this map’s preserved structure help me with the question I’m actually asking? A subway map is wildly geometrically wrong and perfect for catching a train. A photorealistic painting of the subway is far more “accurate” and useless for navigation. Accuracy and usefulness are different axes.

The takeaway: stop grading maps on “how wrong is it?” and start grading them on “is it useful for this, and do I know where it stops being so?” That second clause is the whole next lesson.

When to use it

Keep Box’s line handy in three moments:

• When you’re tempted to dismiss a model for being imperfect. Imperfect is the baseline. The question is whether it’s useful here and whether you know its edges — not whether it’s flawless.
• When you’re tempted to trust a model completely because it’s been “validated” or “data-driven.” Those are still maps. Ask what they were built for and what they therefore drop.
• When you catch yourself saying “that’s just a model” to win an argument. Make yourself name the better map and its track record, or admit you’re just defending your own.

So every map is wrong and some are useful — fine. But when, exactly, does the wrongness stop being harmless and start costing you? That’s not random. It bites in three recognisable situations, and spotting them in advance is the real skill. Next lesson.