Is a Coin Flip Really 50/50? The Physics, the 50.8% Study and What a Digital Coin Does Differently

Everyone learns the same thing at school: a coin flip is 50/50. Two faces, one toss, no argument. It is the standard example of a fair chance, the thing every other probability is compared against. So it was a little embarrassing for the whole idea when a team of 50 researchers spent a year flipping coins, published the numbers, and found that the textbook was slightly wrong.

Not dramatically wrong. Nobody is going to get rich on the difference, and heads is not secretly favored over tails. But the flip does carry a small, measurable, physically explained bias, and knowing what it is (and what it is not) tells you exactly how to flip a coin so that it really is fair. It also explains what a digital coin toss does differently. This is the plain-English version, with every number traced back to a source you can open yourself.

Is a coin flip really 50/50?

No, not exactly. A naturally flipped coin caught in the hand lands on the same side it started roughly 50.8% of the time, so the outcome is very slightly predictable if you know which side was up before the toss. If you do not know the starting side, the flip is, for every practical purpose, a true 50/50.

That is the whole answer, and the rest of the article is the evidence. Three separate claims are folded into it. The coin is not biased toward heads or tails: a normal coin, flipped and caught, comes up heads half the time, so that part of the textbook survives. The coin is slightly biased toward its starting side: start it heads up and it comes back heads up a little more than half the time, and the same goes for tails. And the bias is small, about 0.8 of a percentage point, which takes hundreds of thousands of flips to see and is why nobody noticed for so long.

Why does it matter? Because people use coin flips for real decisions: who kicks off, who gets the last seat, who goes first in a draft. A flip is supposed to be the one method nobody can argue with. As we discuss in why fair decision-making matters, the value of a random method is not just the randomness, it is that everyone involved believes it. A bias of 0.8 points will never change a single flip you can point to, but "the guy who called it was looking at the coin" is exactly the kind of doubt that ruins a fair process.

What did Diaconis, Holmes and Montgomery predict in 2007?

They predicted that a vigorously flipped coin, caught without bouncing, would land the way it started about 51% of the time, because a real toss wobbles as well as spins. Their paper, "Dynamical Bias in the Coin Toss", appeared in SIAM Review in 2007 and set the target that the big 2023 experiment later hit.

Persi Diaconis is a Stanford statistician who was a professional magician before he was a professor, which turns out to be relevant. He and his co-authors, Susan Holmes and Richard Montgomery, started from an observation that sounds like a trick: a coin is not random at all. It is a rigid disc obeying Newton's laws. They built a coin-tossing machine, a spring on a ratchet, and reported that with careful adjustment it landed heads up "one hundred percent of the time" when started heads up. Their blunt conclusion: coin tossing is "physics", not "random". A human hand is not a calibrated spring, of course, and the small differences between one flip and the next are what make the result unpredictable. Earlier work by the mathematician Joseph Keller had shown that if a coin spins cleanly end over end, like a wheel on an axle, a fast flip caught in the hand lands heads exactly half the time.

The wobble that breaks the symmetry

What Diaconis and colleagues added was precession. Flip a coin off your thumb and it almost never spins about a perfect axis in its own plane. It also wobbles, the way a spinning plate wobbles before it settles. In physics terms, the coin's spin axis is not quite lined up with its angular momentum, and the angle between the two stays fixed for the whole flight.

The extreme case makes the effect obvious. Flick a coin so that it goes up spinning flat, like a pizza, and it never turns over. The authors call this a "total cheat coin", and note that magicians and cheats can produce flips that look normal but never actually turn over. Their theorem says that if the wobble angle is under 45 degrees the coin does not flip at all; it only wobbles and always lands as it started.

Real flips are nowhere near that bad, but they all sit somewhere between the perfect end-over-end spin (fair) and the pizza (rigged), and every point in between leans a little toward the starting side. To put a number on it the authors attached a ribbon to a half dollar, filmed flips with a high-speed camera, and measured the wobble on 27 real tosses. The predicted same-side probabilities ranged from 0.500 to 0.545, with a mean of 0.508, which they rounded up and reported as "about .51". They estimated that detecting a bias this small would take about 250,000 tosses, which is why nobody had done it, and they were careful to say the analysis covers coins caught in the hand. A coin that bounces on a table or floor is a messier problem, and we come back to it below.

What did the 350,757-flip study find?

The coin landed on the same side it started 50.8% of the time, with a 95% interval of 50.6% to 50.9%, almost exactly the 51% predicted 16 years earlier. Heads came up 50.0% of the time, so there is no heads-or-tails bias at all. The result won the 2024 Ig Nobel Prize for Probability.

The paper is "Fair coins tend to land on the same side they started: Evidence from 350,757 flips", by František Bartoš of the University of Amsterdam and 49 co-authors. Diaconis had said 250,000 flips would be needed, so they did more than that. Forty-eight people flipped coins of 44 different currencies and denominations, filming as they went, and recorded both the starting side and the landing side of every toss.

The headline numbers, straight from the paper:

The bias belongs to the flipper, not the coin

The most interesting line in the paper is not the 50.8%. It is the spread between people. Some participants landed same-side just under half the time, which is as fair as it gets. One landed same-side about 60% of the time, a lean you could spot in a single afternoon of flipping. The 44 different coins made no difference, which fits the theory: precession comes from the hand, not the metal. The authors also noticed that the bias shrank as people flipped more, and estimate that around 10,000 flips (about 10 hours) might be enough to virtually eliminate it. The training program for becoming a fair flipper is simple and extremely boring.

How much is 0.8 of a point worth? Call 1,000 flips knowing the starting side and you would expect to be right about 508 times instead of 500. Real, but you need a lot of flips to feel it.

In September 2024 the study was awarded the Ig Nobel Prize in Probability. The citation credits Bartoš, Eric-Jan Wagenmakers, Alexandra Sarafoglou, Henrik Godmann and many colleagues "for showing, both in theory and by 350,757 experiments, that when you flip a coin, it tends to land on the same side as it started". A silly-sounding experiment that settled a real 16-year-old question.

Why is the bias about the starting side, not heads versus tails?

Because the physics does not know what is printed on the coin. Precession keeps the coin's "up" face pointing up for slightly more than half its flight, whichever face that happens to be. Start it heads up and it leans heads; start it tails up and it leans tails by the same amount. Averaged over random starting sides, heads comes up exactly half the time.

This is the part people get wrong when the study makes the rounds online. "Scientists prove coins favor heads" is not what was found. Heads came up 50.0% of the time, dead center. What the coin favors is continuity: it comes back the way it went up a little more often than not.

The picture from the 2007 paper is the cleanest way to see it. Imagine a globe where "up" is the north pole and every orientation of the coin is a point on its surface: the northern hemisphere means the starting face is up, the southern means it has turned over. A perfect end-over-end flip traces a great circle through both poles and spends exactly half its time in each hemisphere. A wobbling flip traces a smaller circle tilted toward the starting pole, so it spends slightly more than half its time in the starting hemisphere, and wherever it stops, the starting side is slightly more likely to be up.

Notice what this implies for cheating. The bias only helps you if you know the starting side. A coin flipped by someone else, with the starting face hidden, gives you nothing to work with. That single fact is the basis of every "how to flip fairly" recommendation below.

How do you flip a coin so it really is 50/50?

Do not let anyone see the starting side. That alone removes the exploitable bias, because the bias is entirely about knowing which face was up before the toss. Beyond that, flip high and fast, catch it in the hand, and never decide anything by spinning a coin on a table.

Bartoš put it plainly to the Heidelberg Laureate Forum: the only prerequisite for using coin tosses fairly is that neither party knows which side the coin starts on. The paper says the same thing more formally, that for high-stakes decisions "the starting position of the coin is best concealed". Here is how to do that in practice.

Two things you do not need to worry about. One is the coin itself: the 2007 authors had coins made with lead on one side and balsa wood on the other and found no bias when they were flipped and caught. Weight only matters if the coin is spun. The other is landing on edge. Murray and Teare, cited in the same paper, estimated that a US nickel lands on its edge about 1 flip in 6,000. If it happens, laugh and flip again.

What about streaks? Is five heads in a row suspicious?

No. Five heads in a row from a fair coin has a probability of 1/32, or about 3.1%, so it happens roughly once every 32 five-flip sequences. It is uncommon on any single attempt and completely ordinary over an evening of flipping. The sixth flip is still 50/50.

This is arithmetic, not a study, so you can check it yourself. A fair flip has probability 1/2 of heads. Two heads in a row is 1/2 × 1/2 = 1/4. Each extra flip halves the probability again:

A few more, since people ask: ten heads in a row is 1/1,024, just under 0.1%, and twenty is 1/1,048,576. If you only care that the run is all the same face, double each figure, because five tails is as good as five heads. Five of the same in a row is 1/16.

The gambler's fallacy is the belief that after a run of heads, tails is "due". It is not. The coin has no memory; the 50.8% study is, if anything, a small argument in the other direction, since a same-side lean nudges toward repeating whatever just happened. Whatever came before, the next fair flip is a fresh 1/2. The same instinct makes people think a prize wheel that has not landed on the big-prize slice "must be about to", which we cover in spin wheels for fair prize distribution. The mirror-image error is just as common: seeing a run of five and concluding the coin is rigged. Over a long session, runs are expected. Flip for an hour and never see five in a row, and that would be the surprising outcome.

Which famous decisions came down to a coin toss?

More than you would expect, and the best-documented one is the first airplane flight. On December 14, 1903, Wilbur Wright won a coin toss against his brother for the first attempt at powered flight, and promptly stalled the machine into the sand. When it was repaired three days later it was Orville's turn, and Orville made the first flight in history.

That account comes from the US National Park Service, which runs the Wright Brothers National Memorial at Kitty Hawk. The coin decided who would try first, the physics decided who would succeed, and the history books remember Orville. A coin toss allocates the opportunity, not the outcome.

The NFL coin toss, by the rulebook

Sport is where most people meet a formal coin toss, and the NFL writes its procedure into the rules. Under Rule 4, Section 2, Article 2 of the official playing rules, not more than three minutes before kickoff the Referee tosses a coin at the center of the field in the presence of both teams' captains, and the call of heads or tails must be made by the captain of the visiting team before the toss. The Pro Football Hall of Fame keeps a record of every Super Bowl toss; as of its 2022 tally the Dallas Cowboys had won the most, six.

For years the overtime toss was the most consequential coin flip in American sport, because in the regular season a touchdown on the first possession ended the game. From the 2025 season both teams are guaranteed a possession in regular-season overtime, within a ten-minute period, matching the postseason rule in place since 2022. The toss still matters, but the league deliberately reduced how much a single flip could decide. Given a 50.8% flipper and a captain who can see the coin, that seems wise. It is the same lesson the physicists reached: a toss works because nobody can influence it, and that acceptance is the whole product, as we argue in why spin wheels beat manual decision-making.

How does a digital coin toss work, and is it fairer?

A digital coin toss on Wheel Spin Pro takes 32 random bits from the browser's cryptographically secure random number generator, uses rejection sampling so that neither face is favored by even a rounding error, and only then animates the coin to land on the face already chosen. There is no starting side, no wobble and no hand. The result is 50.000% by construction, and the receipt lets you check that the draw was not tampered with.

The physical flip and the digital flip fail in completely different ways, so it is worth laying them side by side.

What happens when you press Flip

Entropy. Your browser exposes a function called crypto.getRandomValues, fed by the operating system's random source, the same one used to generate encryption keys. It is not Math.random(), which is a fast arithmetic generator that is fine for confetti and wrong for anything with a winner.

Rejection sampling. The engine asks for a 32-bit number, one of 4,294,967,296 values, and turns it into a choice between two faces. For two faces "even or odd" would happen to be exact, but the coin uses the same routine as every picker on the site, where the entrant count might be seven. Dividing 4,294,967,296 by 7 leaves a remainder, so a plain modulo would hand some entrants one extra chance in four billion. Rejection sampling discards any draw above the largest exact multiple of the entrant count and draws again, so every outcome is exactly equally likely.

Winner first, then the animation. The face is decided before the coin moves; the tumble is then choreographed to end on it. An earlier version always started the animation heads-up and rotated a fixed number of half-turns, which made the tumble count a tell. It now starts from wherever the coin is resting and rotates whatever distance is needed, so the animation carries no information beyond the answer itself.

The receipt. Before the draw, the page generates a secret 32-byte seed and shows you its SHA-256 fingerprint. The bytes for the flip are produced by hashing that seed with a client seed and a counter, with the same rejection sampling applied. After the flip, the seed is revealed. Anyone with a generic SHA-256 calculator can check that the revealed seed hashes to the fingerprint shown before the draw, so it could not have been swapped once the labels were known, and that hashing seed, client seed and counter reproduces the bytes that chose the face. Paste the receipt into /verify and the page does both checks in your browser, using the same code that ran the draw. We explain the scheme in plain terms in how spin wheels improve transparency in online communities.

So which is fairer?

On the numbers, the digital coin: no starting side to exploit, no hand to wobble. But the other half of "fair" is whether the people involved believe it, and a real coin has a large advantage there: everybody in the room can see the whole process, with no server and no code to take on trust. That is why the receipt exists. It is the digital coin's answer to "how do I know you did not just pick?"

When should you flip a coin to make a decision?

When the two options are close enough that you have already thought about it more than it deserves. A coin does not know the better answer, but it ends the deliberation, and it gives you a useful second signal: if the result lands and you feel a pang of disappointment, you just learned which option you actually wanted. Use the coin for the lunch order and the window seat, and keep the slow thinking for decisions that are not a toss-up. We go into where randomness stops being helpful in using spin wheels to make decisions faster.

The honest summary

A coin flip is 50/50 in the way that matters and 50.8/49.2 in a way that only a physicist with 350,757 data points can see. The bias is real, it was predicted before it was measured, and it belongs to the flipper's hand rather than to the coin or to either face. It is exploitable only by someone who knows the starting side, which means it is defused by the oldest rule of fair flipping: call it before you look.

If you can hide the starting side and catch the coin, a real coin is a perfectly good decision-maker. If the two parties are apart, or the outcome is worth something, a digital coin with a checkable receipt does the same job with no wobble, no hand and no "were you looking?" argument afterward. Either way, the sixth flip is still 50/50 no matter what the first five did. The coin does not remember, and neither should you.