🪙Is a Coin Flip Really 50/50?

By the WheelJar editorial desk

The coin is fair. The toss is not.

Everyone learns that a coin flip is the definition of an even chance, and for most purposes that is close enough to true. But it is not exactly true, and the reason is genuinely surprising: the bias lives in the throw, not in the metal.

Persi Diaconis, Susan Holmes and Richard Montgomery worked out the mechanics in a 2007 paper in SIAM Review. A coin thrown by a human hand does not tumble neatly end over end. It precesses — the axis it spins about wanders, describing a cone, in the same way a spinning top leans and circles. Because of that wobble, the coin spends marginally more of its flight time showing the face that was up when it left the thumb. Catch it in your hand and that surplus shows up in the result.

Their model put the effect at around 51% for the starting face. Not a huge edge, but a real one, and crucially it is not a property of the coin at all. A perfectly machined, perfectly balanced coin tossed by a person shows the same bias, because the person is the source of it.

350,757 flips later

Physics predictions are cheap. In 2023, a team led by František Bartoš did the tedious part: 48 people flipped coins 350,757 times, across 46 different coin types, recording the starting face and the result of every single toss.

The same-side rate came out at 50.8% — comfortably above 50%, close to the theoretical prediction, and far too large to be chance across a sample that size. It is one of the more satisfying results in recent probability, because it took a two-hundred-year-old assumption, tested it at ruinous scale, and found the textbook slightly wrong in exactly the direction the theory said it would be.

The study also found the effect varied by person. Some flippers showed a strong same-side tendency and some showed almost none, which fits the model: how much a coin precesses depends on how it is released, and people release coins differently.

What 50.8% actually means to you

Very little, and it is worth being clear about that.

The effect amounts to roughly one extra same-side result every 125 tosses. Nobody settling who buys the coffee is being cheated. You would need thousands of flips before the bias emerged from the ordinary noise of a fair coin, which is precisely why it took until 2023 for anyone to demonstrate it directly.

Where it does matter is anywhere the toss is repeated at scale, or where the caller can see the starting face. Concealing the starting position removes the exploitable part entirely. So does letting the coin land and bounce on the floor rather than catching it — the bounce adds enough chaotic energy to wash the precession bias out.

Spinning a coin on a table, incidentally, is a different game altogether and a much worse one. A spun coin falls over onto its rim, so the exact profile and weight distribution of the edge dominate the outcome, and some coins show a heavy skew. If you must use a physical coin, toss it.

Why the digital flip has no same-side bias

The coin flip on this site has no starting face, no wrist and no flight time. Each flip draws bytes from your browser’s cryptographic random number generator — the same source used for encryption keys — and maps them to heads or tails using rejection sampling so that neither outcome is even fractionally favoured.

There is also no memory. The generator does not know what the previous flip was, does not track the running tally, and has no mechanism that could nudge results towards an even split. That is why five heads in a row is unremarkable here: it happens about one time in 32, and the sixth flip is still 50/50. Believing a correction is due is the gambler’s fallacy, and it is the single most common misreading of a fair sequence.

The animation, for the avoidance of doubt, is drawn after the result is chosen. It is theatre, not process — the same arrangement as the wheel spinner, where the winner is picked the instant you press the button and the wheel merely rotates to it.

If you want the underlying detail on what makes one random source better than another, that is the subject of how random is random.