Why a Sealed Box Creates a Decision-Theory Puzzle
The conflict is not really about whether an extra $1,000 is useful. It is about what counts as a reason to choose when your action may reveal what someone—or something—has already predicted.
One way to reason is by dominance: compare the options while holding everything else fixed. Taking both boxes gives you the same contents as taking only the opaque box, plus the guaranteed $1,000. On that comparison, taking both is better.
Another way is to ask what your choice tells you about the box. If the predictor tends to put the million in when it expects a one-box choice, then deciding to take only that box is evidence that the million is likely to be there. Choosing both may be evidence that it is not. This is not a claim that your present choice reaches backward and changes the contents. It is a claim about how strongly your choice is connected to the prediction that determined them.
Two Kinds of “If I Choose This”
Decision theorists distinguish between a choice’s causal effect—what the choice itself changes—and its evidential value—what the choice indicates about other facts. In ordinary situations, these often point in the same direction. Here they pull apart: the boxes are already packed, so taking both cannot add money to the opaque box, but the predictor’s accuracy makes your intended action informative about what is inside.
That split explains why the puzzle resists a simple answer. The two-box argument treats the prediction as fixed and asks which available action yields more. The one-box argument considers how a reliable predictor’s forecast tracks the chooser’s decision-making pattern. Each can look rational under a different account of how decisions should be evaluated.
Why the Puzzle Matters
A predictor does not have to control your decision for its forecast to matter. If it has learned to anticipate people who reason in a particular way, your reasoning style can correlate with the outcome you face. The puzzle therefore exposes a hidden assumption in many decision models: that evaluating an action means considering only what it causes, not what choosing it reveals.
In a real-world setting, the stakes would depend on how reliable the predictor actually is and how its forecasts are linked to your behavior. In the thought experiment, that link is made unusually strong—enough to turn a guaranteed bonus into a test of what “making the best choice” means.
