Artificial intelligence did not first learn to be human by reading books. It learned by playing.
That sounds sentimental, but it is technically accurate. Long before today’s large language models, researchers chose games as laboratories for intelligence: checkers, chess, backgammon, Go, Atari, poker, StarCraft, Dota, Quake, hide-and-seek, robot soccer, table tennis, drone racing, and even the Rubik’s Cube. The list is not accidental. Games are not a childish detour from “real” intelligence. They are one of the cleanest ways to isolate what intelligence does: perceive a situation, choose an action, accept consequences, adapt, and try again.
Arthur Samuel’s checkers program in the 1950s is a good starting point. IBM describes Samuel Checkers and TD-Gammon as early game-playing systems that used play to study strategy and improve through trial and error. Samuel’s work is also closely tied to the very phrase “machine learning”: a machine improving its performance through experience rather than through explicit instruction.
Chess then became the symbolic battlefield. IBM’s Deep Blue defeated Garry Kasparov in 1997, demonstrating that brute-force search, evaluation functions, enormous engineering effort, and domain expertise could defeat a world champion in a narrow but culturally loaded arena. It was not yet the modern learning paradigm, but it showed why games mattered: everyone understood the benchmark. A machine beating a champion at chess was not an abstract metric. It was a public humiliation of human intellectual prestige.
The deeper shift came when machines stopped merely calculating games and started learning them. DeepMind’s 2015 Atari work trained a deep Q-network to play 49 Atari 2600 games directly from sensory input, reaching human-level performance across many of them. This was important because the agent was not given a handcrafted model of Space Invaders or Breakout. It saw pixels, received rewards, and learned policies. A video game screen became a simplified retina; the score became a crude but effective proxy for motivation.
Then came Go. AlphaGo combined supervised learning, reinforcement learning, and search to defeat elite human players. AlphaZero sharpened the point even further: it learned chess, shogi, and Go from self-play, beginning with the rules rather than human games, and reached superhuman strength. The essential trick was almost brutally simple. Let the system play itself. Use winning and losing as the teacher. Replace human instruction with an endless artificial childhood.
This is the first great “why” behind play: games create cheap experience. A human chess player may play thousands of serious games in a lifetime. A machine can play millions, compressing centuries of practice into a training run. The machine does not get bored, humiliated, tired, or protective of its reputation. It explores bad ideas at industrial scale.
The second reason is that games have rules. Reality is messy, but games offer bounded worlds with formal state transitions. That does not make them trivial. Go is simple to describe and enormous to master. Poker adds hidden information. StarCraft adds partial observability, long-term planning, real-time action, imperfect scouting, resource management, and deception. AlphaStar reached Grandmaster level in StarCraft II, while OpenAI Five defeated the Dota 2 world champion Team OG after large-scale self-play training. These were not board games with polite turn-taking; they were chaotic strategic environments with timing, coordination, and opponent modelling.
Poker matters for another reason: it forces machines to act without full knowledge. Libratus defeated top professionals in heads-up no-limit Texas Hold’em, a game in which bluffing, uncertainty, and strategic concealment are central rather than decorative. This is closer to negotiation, cybersecurity, markets, and war than to perfect-information board games.
The third reason is that play generates curricula. In human childhood, play is not random noise. It is self-adjusting difficulty. Children chase, hide, stack, wrestle, imitate, and invent rules. AI researchers rediscovered this in multi-agent systems. OpenAI’s hide-and-seek agents developed escalating strategies and counterstrategies, including forms of tool use the researchers had not explicitly designed into the task. DeepMind’s Quake III Capture the Flag agents learned cooperation with artificial and human teammates. The opponent becomes the teacher; the game becomes an arms race.
Robots add the body back into the story. A chess engine never has to worry about friction. A robot does. That is why robot games are so revealing. RoboCup deliberately uses soccer as a grand challenge for robotics and AI, with the long-term ambition of developing robots that can beat the best human soccer team. Soccer forces perception, locomotion, balance, passing, adversarial tactics, and recovery from failure into one public test.
Table tennis is another almost perfect laboratory. The ball is small, fast, spinning, and unforgiving. Google DeepMind presented a learned robot agent reaching amateur human-level performance in competitive table tennis in 2024. More recently, reports described Sony AI’s table-tennis robot Ace competing with elite players, using reinforcement learning, multiple cameras, and rapid physical control in a dynamic sport. The point is not that ping-pong itself is economically decisive. The point is that the task compresses perception, prediction, motor control, strategy, and adaptation into fractions of a second.
Drone racing does the same thing in the air. The autonomous system Swift combined deep reinforcement learning in simulation with data from the physical world and raced against human champions, winning several races and recording the fastest time. This is play, but it is no toy problem: high-speed navigation under physical constraints is directly relevant to robotics, logistics, inspection, and autonomous systems.
Even the Rubik’s Cube returned as a robotics playground. OpenAI’s Dactyl work trained models in simulation to manipulate a cube with a humanoid robot hand, using automatic domain randomization to bridge the gap between simulated and real physics. The cube was not merely a puzzle. It was a stress test for dexterity, vision, contact dynamics, and robustness.
So the pattern is clear. AI learns through play because play has exactly the properties learning systems need: goals, feedback, repetition, failure, variation, competition, and measurable progress. Games are safe enough to permit stupidity and rich enough to reward strategy. They let machines fail cheaply until failure becomes data.
This also explains why “play” is not the opposite of seriousness. In humans, play is how the nervous system rehearses the world before the stakes become fatal. In machines, it serves a similar role. The board, the screen, the simulator, the court, the drone gate, the robot soccer field: these are artificial childhoods.
The unsettling conclusion is not that machines are becoming human because they play. It is that we may have underestimated play. Play was never a decorative human pastime. It was always a training regime for intelligence. We taught machines to play because play was the one human method we could formalize. Then the machines did what children do, only faster: they played until they became dangerous.
Leave a Reply