← All articles
mathprobabilityrandomnessphysicsSeptember 17, 20264 min read

What Is a Coin Toss? Not Actually Random, and Not Actually Fair

By the BrainSnail editorial team. How these articles are written and checked, and how to tell us when one is wrong.

Flipping a coin is the standard image of a fair fifty-fifty choice. The flip is a deterministic physical process, a machine can make it land the same way every time, and careful measurement of human flips finds a small but real bias that theory predicted in advance.

Determinism dressed as chance

A tossed coin obeys ordinary mechanics, so its outcome is fully determined by the initial conditions: the force applied, the point of contact, the spin rate, the angle, the height and the surface it lands on. Nothing genuinely random enters. What makes it a usable randomiser is sensitivity, since a change in initial conditions far too small for a person to control changes the outcome, so the result is unpredictable in practice while being determined in principle. That is the same property that makes dice and roulette usable. Its status as a randomiser is therefore a statement about human motor control rather than about physics, and it fails when the sensitivity is removed. Persi Diaconis and colleagues built a mechanical flipper that controls the initial conditions precisely and produces the same result every time, which demonstrates the point directly, and trained people can bias outcomes measurably.

The measured biases

Several distinct effects have been identified and quantified:

  • A same-side bias, predicted by physical analysis showing that a coin precesses and therefore spends slightly more time with its initial face up, so it lands the same way it started slightly more often than half the time
  • A large collaborative study published in 2023, involving hundreds of thousands of recorded flips by many people, which found that bias at about fifty-one percent, matching the prediction
  • Substantial variation between individuals, with some flippers showing much stronger bias and others none, and with practice reducing it
  • Spinning a coin on a table rather than flipping it, which produces much larger biases because the coin's edge is not perfectly symmetrical and it tends to fall towards the heavier face
  • Catching versus letting it land, which changes the physics and the bias
  • Balance imperfections in real coins, which matter far less for flipping than the popular belief suggests, since a flipped coin's rotation averages them out

Why it still works

A one percent bias is negligible for nearly every practical purpose, which is why coin flipping remains a perfectly reasonable way to settle an argument, choose who kicks off or allocate a seat. The relevant question is always whether the bias is large relative to what is at stake and whether either party can exploit it. Both conditions fail in ordinary use, since neither party knows which face started upward in a way they could act on and the effect is too small to matter over one trial. They do not fail in every case. Where the same person flips repeatedly, where large sums depend on it, or where a participant has practised, the assumption of fairness becomes questionable, and the standard protections are to have one party flip and the other call, to call in the air, and to require the coin to land rather than be caught. Those conventions predate the physics and neutralise exactly the effects it identifies.

Where genuine randomness comes from

Applications that need randomness they can rely on do not use coins. Computers use pseudorandom generators, which are deterministic algorithms producing sequences that pass statistical tests, and which are entirely predictable to anyone who knows the internal state, which is why they are unsuitable for cryptography without additional measures. Cryptographic applications use hardware sources drawing on physical processes including thermal noise in circuits, timing jitter and radioactive decay, and operating systems combine several such sources. Quantum processes are the only physically fundamental source, since quantum measurement outcomes are genuinely indeterminate under the standard interpretation rather than merely unpredictable, and commercial quantum random number generators exist. Lotteries use elaborate mechanical draws with public observation and independent auditing, which addresses trust as much as randomness. The general principle across all of them is that a randomiser must be unpredictable to the specific people who could benefit from predicting it, which is a security requirement rather than a mathematical one.

The takeaway

The flip is fully determined by its initial conditions and works only because it is too sensitive for a person to control, which a mechanical flipper defeats entirely. Physical analysis predicts a same-side bias of about fifty-one percent, and a study of hundreds of thousands of flips found exactly that. The bias is far too small to matter in ordinary use, and calling in the air while somebody else flips neutralises it anyway.

Practise this

Questions from Probability and Statistics

Reading about something is not the same as being able to recall it. These are real questions from the Probability and Statistics unit in our Math track, answers and explanations included. The unit has 120 in total across 21 steps.

  • Choose all that applyLevel 3

    1. Which of these describe a good sample? Select all that apply.

    • It should be chosen randomlycorrect
    • It should represent the populationcorrect
    • It must include the whole population
    • More bias makes it better

    Good samples are chosen randomly and represent the population; a sample is not the whole population, and less bias is better.

  • Guess the numberLevel 4

    2. Find the range of this data set: 12, 19, 25, 40, 47.

    Answer: 35

    Range = 47 - 12 = 35.

  • Match the pairsLevel 3

    3. For the data set 2, 2, 3, 5, 8, match each measure to its value.

    Answer: Mean = 4; Median = 3; Mode = 2

    The mean is 20/5 = 4, the middle value is 3, and 2 appears most often.