Fallacy of Composition

The fallacy of composition is a specific reasoning error. One observes something about individual parts and claims the same holds for the whole. An example: Every player on a team is an excellent goal scorer. So, the conclusion goes, the team is excellent. That need not be true, because no one defends and no one passes. The error doesn’t lie in the observations, but in the transfer. The whole often has properties that no single part has.

Why the conclusion is so tempting

The reasoning error seems harmless because it sometimes works. If every brick in a house is red, the house is red. With color, the transfer works fine; with performance, cost, or safety, almost never. That’s why, when reading, one rarely notices when the line has been crossed.

In economics, the fallacy is costly. A single saver who saves more becomes wealthier. If everyone saves more at the same time, spending falls and so do incomes. Economists call this the paradox of thrift. Many correct individual statements turn into one false overall statement.

Skepticism is also worthwhile with tech news. That every component of a system has been tested doesn’t mean the system has been tested. Errors usually arise in the interplay, not in the individual part. That’s exactly where fallacies of composition fail to look.

How to recognize the leap from part to whole

One checks such a claim in three steps. First: Which property is being claimed? Second: Does it hold for the parts or for the whole? Third: Can this property even simply be added up? It’s exactly the third step that’s decisive.

Some properties are additive. Weight is one example: the total weight is the sum of the individual weights. Other properties only emerge from the relationships between the parts. Reliability, speed, or comprehensibility belong to this category. For these, the inference from part to whole is fundamentally invalid.

There’s also the counterpart, the fallacy of division. There, one transfers a property of the whole onto every part. For instance: The corporation is profitable, so every department is profitable. Both errors are mirror images of the same problem. And both are not lies but invalid inferences: the premises can be perfectly true.

The fallacy in AI debates and stock market reports

In AI discussions, this reasoning error is commonplace. A language model consists of simple computational steps that, individually, understand nothing. But it doesn’t follow that the overall system can’t do anything. The reverse holds just as well: because the model performs well on ten individual tests, it isn’t automatically suited for continuous operation.

Something similar happens with safety claims. Providers like to show that every component of a system has been checked. The interesting failures arise where several components react to one another. One example is trading algorithms on the stock market: each one individually behaves reasonably, but together they amplify a market crash.

Market forecasts are also susceptible. If each of twenty startups can plausibly grow, that doesn’t imply growth for the entire industry, because they compete for the same customers. Anyone reading such reports should ask one question: Is a property being summed here that can’t actually be summed? That question usually exposes the fallacy of composition immediately.