In September 1990, the magazine Parade published a column by Marilyn vos Savant that provoked what is arguably the most embarrassing episode in the history of popular mathematics. Over the following months she received more than 10,000 letters, roughly a thousand of them from people with PhDs, telling her she was wrong.

She wasn’t. And the debate she provoked — now called the Monty Hall problem — is one of the cleanest demonstrations of how unreliable human intuition is when probability involves conditional information.

The setup

A game-show host stands in front of three closed doors. Behind one door is a car. Behind the other two are goats. The host knows where the car is; you don’t.

You pick a door — say, Door 1. The host, who knows what’s behind each door, now opens one of the other two doors to reveal a goat. Say he opens Door 3. He then asks: do you want to stay with Door 1, or switch to Door 2?

The overwhelming majority of people say it doesn’t matter: two doors are left, the car is behind one of them, so the probability is 50/50 either way.

That answer is wrong. You should switch. Switching wins 2/3 of the time.

The quickest intuition

The easiest way to convince yourself is to enlarge the problem. Imagine 100 doors instead of three. You pick one — the chance you picked the car is 1/100. The host then opens 98 other doors, all goats, leaving just the door you picked and one other.

Now does it feel like 50/50?

Of course not. You picked essentially at random from 100 doors; the host then carefully avoided the car 98 times in a row. The remaining closed door is overwhelmingly the car. The same logic holds at three doors — just less dramatically. Your original pick was right with probability 1/3. The host’s action concentrates the remaining 2/3 into the one unopened door.

The Bayesian version

For readers who want the formal version: let AiA_i be the event “the car is behind door ii,” and let HjH_j be the event “the host opens door jj.”

You pick Door 1. The host opens Door 3. We want:

P(A1H3) versus P(A2H3)P(A_1 \mid H_3) \text{ versus } P(A_2 \mid H_3)

Bayes’ theorem gives:

P(AiH3)=P(H3Ai)P(Ai)P(H3)P(A_i \mid H_3) = \frac{P(H_3 \mid A_i) \cdot P(A_i)}{P(H_3)}

The priors are equal: P(A1)=P(A2)=P(A3)=1/3P(A_1) = P(A_2) = P(A_3) = 1/3.

The key numbers are the likelihoods:

  • P(H3A1)=1/2P(H_3 \mid A_1) = 1/2 — if the car is behind your door, the host chooses freely between Doors 2 and 3.
  • P(H3A2)=1P(H_3 \mid A_2) = 1 — if the car is behind Door 2, the host must open Door 3 (he won’t reveal the car, and he won’t open your door).
  • P(H3A3)=0P(H_3 \mid A_3) = 0 — the host never opens the door hiding the car.

Plugging in gives P(A1H3)=1/3P(A_1 \mid H_3) = 1/3 and P(A2H3)=2/3P(A_2 \mid H_3) = 2/3. Switching wins twice as often.

Why our intuition fails

The Monty Hall problem is a masterpiece of cognitive trickery because it hits three biases simultaneously.

Equiprobability bias. When presented with nn possibilities, our first instinct is to assign each the probability 1/n1/n. That’s correct only when the underlying process is symmetric — which, after the host acts, it isn’t.

Insensitivity to the information-generating process. Most people forget that the host’s action is not random. He has information you don’t, and his behavior depends on where the car actually is. If he opened doors randomly, the problem would indeed be 50/50 after he happened to miss the car. The non-randomness of the host is the entire crux.

Anchoring on the original choice. Psychologically, people feel their first choice “deserves” to stay correct. Switching feels like admitting error. The math doesn’t care.

Computational demo

If you still don’t believe it, simulate it. A 20-line Python script does the trick:

import random
wins_switch = 0
wins_stay = 0
trials = 100_000
for _ in range(trials):
    doors = [0, 0, 0]
    doors[random.randint(0, 2)] = 1  # place the car
    pick = random.randint(0, 2)
    # host opens a door that is neither the pick nor the car
    options = [i for i in range(3) if i != pick and doors[i] == 0]
    opened = random.choice(options)
    # the door you'd switch to
    switch_to = [i for i in range(3) if i != pick and i != opened][0]
    if doors[pick] == 1: wins_stay += 1
    if doors[switch_to] == 1: wins_switch += 1
print(wins_stay / trials, wins_switch / trials)

You’ll see numbers very close to 0.333 and 0.667. No amount of intuition will override that, but a few thousand simulated trials will.

The broader lesson

The Monty Hall problem is not really about game shows. It’s about what mathematicians call conditional probability — the idea that the probability of an event depends on what else you know.

Most real-world probability problems are conditional. A medical test comes back positive; what’s the probability you’re sick? A neural network flags an email as spam; what’s the probability it actually is? The same kind of reasoning that trips people up on Monty Hall trips people up on medical screening, legal evidence, and machine learning. The common mistake is ignoring the process that generated the information.

Probability is not a statement about the world; it’s a statement about the world given what you know about it. Update carefully.

Frequently asked

Is the Monty Hall solution still 2/3 if the host opens a random door?

No. If the host chooses randomly and happens to reveal a goat, your winning probability after switching is only 1/2 — the same as staying. The extra information comes specifically from the host knowing where the prize is.

Why did so many professional mathematicians get this wrong?

Because the problem is designed to trigger a familiar mental shortcut: 'two doors left, so 50/50.' Most probability intuition is trained on independent events, and the Monty Hall setup is anything but independent — the host's action is conditioned on your first choice.