Seventy years of games in which the human side stopped being the strongest
How to read this
Every entry is a moment when a game changed hands. Red dots are the machine's milestones, green ones are human wins worth remembering — there are fewer of them, and they are getting rarer.
The pattern repeats: a game is declared out of reach for decades, then falls in a single match, and within a year or two the result stops surprising anyone. The interesting part is not the falling but how quickly each defeat becomes normal.
1951
The first game programsboard
Christopher Strachey writes a draughts program for the Ferranti Mark I; Dietrich Prinz writes one that solves mate-in-two chess problems. It cannot play a whole game — it tries every move in turn and takes minutes doing it.
1956
Samuel teaches a machine to learnboard
Arthur Samuel runs a checkers program on the IBM 701 that tunes its own evaluation from the games it plays — one of the first working demonstrations of machine learning, and it beats its author.
1979
Backgammon: BKG 9.8 beats the world championdice
Hans Berliner's program beats world champion Luigi Villa 7-1 in Monte Carlo — the first time a reigning world champion loses to a machine at anything.
Berliner himself said the dice ran the program's way that match.
1988
The first grandmasterchess
Deep Thought beats Bent Larsen in a tournament game — the first time a computer defeats a grandmaster at a normal time control.
1994
Checkers: Chinook and Marion Tinsleyboard
The world championship match between Chinook and Marion Tinsley — a man who had lost a handful of games in forty years. After six draws Tinsley withdraws for health reasons and the title goes to the program.
Tinsley died of cancer months later; the match was never finished.
1996
Kasparov wins the first matchchess
Deep Blue takes game one — the first time a computer beats a reigning world champion at classical time control. Kasparov still wins the match 4-2 once he works out what the machine does not understand.
1997
Deep Blue beats Kasparov, 3½-2½chess
11 May, New York. Six games; the decisive one lasts nineteen moves. Still the symbol of the whole subject: the machine beat the strongest chess player alive.
Deep Blue searched about 200 million positions per second and understood nothing beyond what was programmed into it.
2007
Checkers is solvedboard
Jonathan Schaeffer's team announces that English draughts is a draw with perfect play. Eighteen years of computation. From that day the game is closed — not for humans, but for everyone.
2011
Watson wins Jeopardy!language
IBM Watson beats Ken Jennings and Brad Rutter, the two strongest players the quiz show ever had — the first machine win in a contest built on language: puns, hints and general knowledge.
Jennings signed his final answer: "I for one welcome our new computer overlords."
2016
AlphaGo beats Lee Sedol, 4-1go
Go was supposed to be humanity's reserve for another twenty years: more positions than atoms in the observable universe, and brute force does not work. AlphaGo wins in March against one of the strongest players in the world.
Move 37 in game two was first called a mistake by professionals. It is now studied as a discovery.
2016
Game four: move 78go
Lee Sedol finds a move AlphaGo did not expect and wins game four. It remains the last human win over a top Go program on even terms.
2017
Poker: Libratus against four professionalscards
Twenty days, a hundred and twenty thousand hands of heads-up no-limit hold'em, and a decisive win for the program. This is a game of incomplete information: bluffing is not decoration here, it is part of the optimal strategy.
2017
AlphaGo Zero: not a single human gamego
A version trained only by playing itself beats its predecessor 100-0. Two months later AlphaZero learns chess and shogi the same way in hours. Human experience turned out not just unnecessary but a hindrance.
2019
StarCraft II and Dota 2video games
AlphaStar beats professionals at StarCraft II; OpenAI Five beats world champions OG at Dota 2. Incomplete information, real time and team coordination — the things that were supposed to stay human.
Both results came with caveats: limits on action rate, camera view and the pool of characters.
2019
Debate: the human still winslanguage
IBM's Project Debater argues live against world champion Harish Natarajan and loses on the audience vote. The machine collects arguments; it does not persuade.
2019
Pluribus: six-player pokercards
Heads-up is still a two-sided game with an equilibrium. Pluribus beats five professionals at once, and multiplayer games have no such tidy theory to lean on.
2022
Gran Turismo: racing at the limitvideo games
Sophy beats the best Gran Turismo Sport drivers. Reaction speed is not the whole story: it races wheel to wheel without breaking racing etiquette.
2022
Cicero and Diplomacylanguage
A game won by negotiation: you have to bargain, promise and betray at the right moment. Cicero plays an online league in natural language and lands in the top ten percent of players.
2023
A weakness humans foundgo
Researchers find a strategy top Go programs do not understand: encircling a large group along the edge. Amateur Kellin Pelrine wins fourteen of fifteen games against a top bot playing that plan, without computer help during the games.
Superhuman turned out not to be uniform: it has holes, humans just do not find them unaided.
What this list does not say
It does not say machines are better at everything. Every line here is a task with clear rules, a clear score and an unlimited supply of practice games. That is exactly the shape of problem where a machine has every advantage and a person has none.
And each victory settles one game, not the species. Checkers is solved; you can still play it with your grandson. Deep Blue won in 1997, and chess is more popular now than it was then — because the point of playing was never to be the strongest thing on the planet.