Strip away the hype and the current wave of AI does one thing: it predicts. Chapter 3 is a reset on what that actually means and why it matters economically.
Prediction is filling in missing information. The authors define prediction as the process of taking information you have and using it to generate information you need but don’t have. A bank has a record of millions of past transactions and whether each turned out to be fraudulent. It does not have the answer for the transaction happening right now. Prediction bridges that gap. The same logic covers medical diagnosis (symptoms in, likely condition out), translation (a sentence in one language in, its counterpart out), and driving (what the sensors see in, what the car should do next out). What looks like a dozen different capabilities is one capability pointed in different directions.
Cheap prediction means more prediction. This is where the economics comes in. Prediction is an input to decision-making, and the recent advances in AI have made that input dramatically cheaper. When the cost of an input falls, we use more of it, and we start using it in places where it was previously too expensive to bother. Nobody used to “predict” what a sentence in Mandarin meant; they hired a translator or did without. Now prediction is cheap enough to deploy on every email and menu. The authors’ earlier book, Prediction Machines, was built on this one idea, and it still holds: as prediction gets cheaper, AI will spread.
Substitutes fall, complements rise. The second economic consequence is about what happens around prediction. When something gets cheaper, the things that do the same job become less valuable, and the things that pair with it become more valuable. The main substitute for machine prediction is human prediction, so the value of a person whose job is mostly forecasting, classifying, or estimating will decline. The main complements are two: data and judgment. Data is what the machines learn from, so organizations that have it, or can collect it, are sitting on an appreciating asset. Judgment is what you combine with a prediction to actually decide something.
Prediction is likelihood; judgment is desire. This distinction is the heart of the chapter. A prediction tells you how likely each outcome is. Judgment tells you how much you care about each outcome. Every decision is the two together: you weigh what might happen against what you want to happen. The machine can tell you there is a 3 percent chance this transaction is fraudulent. It cannot tell you whether blocking a legitimate customer is worse than letting a fraudulent one through, or by how much. That’s a question of values, trade-offs, and context, and it stays with humans. As prediction becomes abundant, the scarce and valuable skill is knowing what you want.
Correlation is not causation, and AI doesn’t know the difference. The authors flag what may be the most common misuse of AI predictions: treating the patterns a model finds as if they explain anything. Models find correlations. For many purposes that’s enough; if the goal is to guess which customers will churn, you don’t need to know why. But if you want to act on the world, to change the thing rather than predict it, correlation can lead you badly astray. A model might learn that customers who call support are more likely to leave, but that doesn’t mean cutting the support line will keep them. When the question is causal, the right tool is not a bigger model but a randomized experiment, which remains the most reliable method statisticians have for discovering what causes what. Good AI practice involves knowing which kind of question you’re asking.
The Amazon thought experiment, revisited. In Prediction Machines, the authors ran a thought experiment about Amazon’s recommendation engine. Imagine it getting steadily more accurate. At first, the effect is modest: better recommendations, slightly more sales. But keep turning the dial and at some point a different question becomes possible: if we’re this good at predicting what people want, why wait for them to order it? Just ship it. Shopping would flip from “order, then ship” to “ship, then keep or return.”
Amazon did file a patent on something like this, which the authors call anticipatory shipping. It has not adopted the model. Chapter 3 asks why, and the answer is the book’s thesis in miniature. Better recommendations on the existing platform are a point solution; they ride on the Amazon system exactly as it is. Shipping before ordering would be a system solution, and the system it would strain most is returns. Amazon’s returns process is so costly that in many cases it’s cheaper to discard a returned item than to restock it. A business model built on sending people things they didn’t ask for, and taking back a meaningful fraction of them, would collapse under that cost structure. To make it work, Amazon would have to redesign logistics, returns handling, inventory, and pricing all at once.
The authors are candid that this is what they got wrong the first time. The threshold in their thought experiment wasn’t a matter of accuracy crossing some number. It was the point where capturing the value of the prediction required leaving the point solution behind and rebuilding the system around it. Prediction Machines saw the dial; Power and Prediction is about what’s on the other side of it.
