N noduly critical thinking · probability puzzles
🔥0 · 0/5 All modules

Probability Intuition Puzzles

Human intuition is reliably bad at probability. The classic puzzles aren't tricks — they're places where the obvious answer is wrong, and the correct answer is teachable.

Why intuition fails here

We evolved to make quick decisions about predators and food, not to combine percentages. So we substitute easier judgments: "feels likely," "I've seen this before," "this story makes sense."

The fix isn't more intuition — it's the math, or a clear table that does the math for you. The puzzles below are the ones you can stop being wrong about for the rest of your life.

The toolkit you need

  • Independence. Each coin flip doesn't remember the last.
  • Base rates. Always start with how common the category is in the population.
  • Conjunctions. P(A and B) ≤ both P(A) and P(B). Adding details lowers probability.
  • Complement trick. Often easier to compute "at least one" as 1 − P(none).
  • Bayes' rule. P(H|E) = P(E|H)·P(H) / P(E). It rules them all.

Puzzle #1 — The Monty Hall problem

Three doors, one car. Should you switch?

A game show: three doors, one hides a car, two hide goats. You pick a door. The host (who knows where the car is) opens a different door, always revealing a goat, and offers you a switch. The intuitive answer is "50/50, it doesn't matter." The correct answer is switch — you win 2/3 of the time. Don't just read about it; run it.

Pick a door.
Stay strategy
0/0
Switch strategy
0/0
Why it works: when you first pick, P(car behind your door) = 1/3. The host's reveal doesn't change that — your door is still 1/3. The remaining unopened door now carries the other 2/3, because the host's information narrows where the car can be.

Puzzle #2 — The Birthday Paradox

How many people before a shared birthday is more likely than not?

A room of 23 people. What's the chance at least two share a birthday? Most people guess 5–10%. The actual answer is just over 50%. With 50 people, it's 97%. With 70, it's effectively certain.

P(any shared birthday)
P(no two share)
Pairs to compare
Why it's surprising: it's not asking "match my birthday" — it's asking "any match in the room." With 23 people, that's 253 distinct pairs, each with about a 1/365 chance to share. The probability of zero matches across all 253 pairs is small.

Puzzle #3 — Base rates and Bayes' rule

You test positive for a 1%-prevalence disease. Should you panic?

Adjust prevalence, sensitivity, and specificity. Watch P(disease | positive test) — it's almost never what you think.

True positives
False positives
P(disease | positive)
Per 1,000 tested
The lesson: Bayes' rule says P(H|E) = P(E|H)·P(H) / P(E). A "95% accurate" test on a 1% disease still leaves most positives false, because the well group is 99× larger than the sick group.

Puzzle #4 — Gambler's fallacy

"Tails is due" is always wrong

A fair coin lands heads seven times in a row. What's the probability of tails on the next flip? 50%. Independent events don't keep score. Click "flip 10" to see streaks emerge naturally.

Total flips
0
Heads
0
Tails
0
Longest streak
0
Why streaks are normal: in 100 fair flips, a run of 6 or 7 of the same side happens almost always. The streak is not evidence the coin "remembers" — it's evidence that runs of independent identical outcomes are common.

Puzzle #5 — The Linda problem (conjunction)

"Linda is 31, single, outspoken, philosophy major, marched for civil rights…"

Which is more probable?

A) Linda is a bank teller.
B) Linda is a bank teller and active in the feminist movement.

In Tversky & Kahneman's 1983 study, ~85% of subjects picked B. It's wrong. P(A and B) can never exceed P(A) — adding a detail can only narrow a set, not widen it.

The trap: the second option matches the story better, so it feels more probable. Representativeness is not probability.

Puzzle #6 — Two-child paradox

A family has two children. At least one is a boy. What's P(both are boys)?

Intuition: 1/2. Correct answer: 1/3.

The sample space is {BB, BG, GB, GG} — all four equally likely. "At least one boy" rules out GG, leaving {BB, BG, GB}. Of those three, only BB has two boys. So P = 1/3.

If the question is "the older one is a boy, what's P(both are boys)?" the answer flips to 1/2 because we've fixed the older child. How the information arrives changes the conditional.

Puzzle #7 — Expected value

Why most lotteries are a bad bet — and some are mathematically excellent

Expected value (EV) = Σ (probability × payoff). Powerball jackpots reach billions, but ticket EV is typically far below the $2 price after taxes and split-jackpot risk. State-lottery scratchers run ~50–70 cents on the dollar in EV.

Same math works for any bet: positive-EV bets are rational to take repeatedly; negative-EV bets aren't. Loss aversion makes humans refuse +EV bets when the loss feels big — see cognitive biases for the corrective.

Puzzle #8 — Live Bayesian updater

Watch a belief update with evidence

A coin is one of two types: fair (P(H) = 0.5) or biased (P(H) = 0.8). You flip it and get the result below. Use the slider to set your prior probability that the coin is biased, then watch how each flip changes your posterior via Bayes' rule.

Flips seen
Current P(biased | evidence)

Why it works: P(biased | flip) ∝ P(flip | biased) · P(biased). Heads boosts the biased hypothesis (0.8 vs 0.5); tails downweights it (0.2 vs 0.5). Stack enough flips and the posterior approaches certainty.

Puzzle #9 — Sleeping Beauty

"What's the probability the coin landed heads?"

Sleeping Beauty is told the experiment: on Sunday a fair coin will be flipped. If it lands Heads, she's woken once (Monday) and her memory wiped. If it lands Tails, she's woken twice (Monday and Tuesday), her memory wiped between. Now she's just been woken. Asked: "what's the credence the coin landed Heads?"

Why it's a real open problem: philosophers and probabilists actually disagree. The halfer respects the prior on the coin; the thirder uses anthropic indexing (you're picking randomly from awakenings, not from coin flips). Neither answer is mathematically wrong — they're answering subtly different questions, which is itself the lesson.

Puzzle #10 — Newcomb's problem

One-box or two-box?

A near-perfect predictor offers you two boxes. Box A is transparent and contains $1,000. Box B is opaque and contains either $1,000,000 or $0. The predictor decides the contents in advance:

  • If they predicted you would take only B, they put $1M in B.
  • If they predicted you would take both A and B, they put $0 in B.

You know all this. Predictor's track record: 1,000 previous players, 999 predictions correct. Do you one-box (take only B) or two-box (take both)?

Why it splits philosophers: Evidential decision theory says one-box (your choice is evidence about your nature, which the predictor saw). Causal decision theory says two-box (your choice now can't cause the past prediction). Both schools have defenders; the disagreement persists because rationality itself is being defined differently.

Connect the dots

Quiz

15 questions on probability puzzles.

0
Score
0
Streak
1/15
Question

Flashcards

Twenty essential probability concepts.

1 / 20
Mastery: —
Space flip · J/ next · K/ prev · 1/2 grade
Daily Probability Puzzle
One fresh counter-intuitive puzzle every day.

Teacher mode

Lesson outline, quick-reference card, and a printable worksheet with answer key.

Lesson outline (45 min)

  • 5 min · Hook — Run a vote on the Monty Hall problem. Most of the class will say "no difference." Run the simulator — switching wins ~67%.
  • 10 min · Concept — Independence, base rates, conjunctions, complement trick, Bayes' rule. Why intuition fails on each.
  • 10 min · Birthday paradox — Walk the formula (1 − 365·364·…·(365−n+1)/365ⁿ). Show n=23 → 50%, n=50 → 97%, n=70 → 99.9%.
  • 15 min · Bayes practice — On the board: P(disease) = 1%, sensitivity 99%, specificity 99%. Have students compute P(disease|+) from the table before the teacher reveals ~50%. With 0.1% prevalence: ~9%.
  • 5 min · Wrap — Homework: name one decision in their life where they used base-rate-free intuition. Compute the base rate.

Quick reference — formulas

Bayes' rule
P(H|E) = P(E|H)·P(H) / P(E)
Posterior = likelihood × prior, normalized.
Independence
P(A and B) = P(A)·P(B)
Only when A and B are independent. Otherwise use conditionals.
Complement
P(at least one) = 1 − P(none)
Birthday paradox, "rolled at least one six," all use this.
Conditional
P(A|B) = P(A and B) / P(B)
"Given B, how likely is A?" Always anchor in the conditioning event.
Expected value
EV = Σ pᵢ · xᵢ
Long-run average payoff. Positive EV → rational to repeat.
Birthday
1 − ∏(365−i)/365 for i=0…n−1
P(any match in n people).

Worksheet