Yes, a robot can beat you at Rock Paper Scissors every time, but only by reacting faster than you can see, not by predicting you. The famous example is a janken robot built at the Ishikawa Oku Laboratory at the University of Tokyo, first shown in 2012, which reads the shape of your hand with high-speed vision and forms the winning shape within about a millisecond. Software that genuinely predicts people is a different thing, and it can beat most humans over many rounds, though not every round.
The robot that always wins
The Tokyo robot uses a camera running at very high speed to track a human hand as it moves. Recognition of the hand's position and shape takes about one millisecond, and a fast robotic hand then makes whichever shape beats it. To a human watching, both hands appear to land at the same moment. In reality the robot has already seen your throw.
The lab followed with improved versions. A second version finished its hand shape almost exactly as the human's did, making the reaction even harder to notice, and a third added fast pan-tilt tracking so the system could follow a hand across a wider field of view.
In a refereed match, this would be a late throw, and it would lose the round. It is a brilliant demonstration of high-speed vision and control, which was the point of the research, but it is not a strategy. You cannot learn anything from it about how to play.
Machines that predict instead of react
The more interesting question is whether a machine can beat you without seeing your hand early. It can, if you are predictable, and almost everyone is.
- The RoShamBo programming competitions. In 1999, Darse Billings organised the First International RoShamBo Programming Competition, where programs played long matches against each other. The winner, Iocaine Powder by Dan Egnor, combined many prediction strategies with "meta-strategies" that assumed the opponent was trying to predict it back.
- The New York Times game. The newspaper published an online game in which the computer used a record of over 200,000 previous rounds to guess what a human would throw next, based on recent moves.
- Kaggle, 2020. Kaggle ran a Rock Paper Scissors simulation competition in which bots played matches of 1,000 rounds, long enough for pattern detection to matter.
The common method is simple to describe: look at the last few throws and results, find what this player (or players in general) usually did in the same situation, and play the counter. Research like Wang, Xu and Zhou (2014), which found win-stay and lose-shift tendencies in human play, explains why this works. More on that is on our science page.
Can AI beat a random player?
No. Against a player who throws each option with equal probability and no pattern, every strategy, human or machine, wins about one third of the time over the long run. Game theory says that random play cannot be exploited, which is why predictive AI only has an edge over people, not over dice. Our game theory page goes through the reasoning.
Try it yourself
When you play against the computer on WRPSA, the computer chooses its throw from completed rounds only and cannot see your current throw. That makes it a fair test of whether you are as random as you think. Our own data from 3,020 throws suggests most people are not: rock was thrown 36.2% of the time and scissors only 28.2%, in a sample mostly of people playing the computer between July and September 2026.
FAQ
Is there a robot that always wins Rock Paper Scissors?
Yes, the University of Tokyo janken robot, but it wins by seeing your hand and reacting in about a millisecond.
Can machine learning improve Rock Paper Scissors strategy?
It can find and exploit patterns in a person's play. It cannot beat a truly random opponent.
Is the robot cheating?
By tournament rules, yes: reacting to the opponent's hand is a late throw. See our rules.