Everybody remembers the rule: in the Monty Hall problem, switch doors.
The rule works. It also comes with a job description for the host that tends to disappear during retellings. The host knows where the car is, always opens a door hiding a goat, and always offers the switch. Remove the first qualification and the famous two-thirds advantage vanishes.
I simulated one million games with the usual informed host, then another million with a host who did not know where the car was. The second host simply chose one of the player's two unselected doors at random. When that door exposed the car, the round ended in a small but decisive failure of game-show production.
informed host guessing host
trials 1,000,000 1,000,000
goat revealed 1,000,000 665,805
car accidentally revealed 0 334,195
among rounds with a goat:
stay wins 33.244% 49.931%
switch wins 66.756% 50.069%
The informed host produced the familiar result. The guessing host turned switching into a coin toss.
What the host knows changes what the goat means
The puzzle traces to a pair of 1975 letters by statistician Steve Selvin. In his original setup, the host's behavior is part of the mathematics: he knows where the prize is and avoids revealing it. A later analysis by statistician Richard Gill argues that the problem is really a lesson in mathematical modeling, because apparently minor assumptions about the host define which game is being played.
With an informed host, your first choice is right one-third of the time. If it is wrong, which happens two-thirds of the time, the host is forced to open the only other goat door. Switching collects all of those initially wrong guesses. The host does not improve your first choice; he uses private information to remove the losing alternative beside it.
A guessing host creates a different sample of games. Suppose your first choice is the car. Both unselected doors hide goats, so the round always continues. Now suppose your first choice is a goat. One unselected door hides the car and the other hides the second goat, so the host has only a fifty-fifty chance of keeping the game alive.
That leaves two equally likely routes to the moment when you see a goat:
- You chose the car initially: probability one-third, and the host certainly reveals a goat.
- You chose a goat initially: probability two-thirds, but the host reveals the other goat only half the time. That also contributes one-third.
Once we condition on the rounds that reached the offer, the initial car and the remaining car are equally likely. The simulation's 50.069% switching result is ordinary random variation around one-half, not a tiny surviving advantage.
The experiment has a narrow doorframe
I fixed the player's first choice, placed the car uniformly at random, and used Python 3.14.9 with a fixed seed. Fixing the player's door does not affect the result because the three doors are symmetric. The simulation checks the arithmetic for these two host protocols; it does not model the television show, a host with a strategy, or a host who sometimes withholds the offer.
The useful part of Monty Hall is not merely "switch." An action can carry information about knowledge you cannot see. A goat revealed by someone who was forbidden to reveal the car is evidence; the same goat revealed by someone guessing only tells you that the game survived long enough for you to be asked.