Self-Play

For an AI system to get better, it needs opponents or conversation partners who challenge it. Humans can take on this role — but humans are slow, expensive, and at some point no longer good enough. Self-play solves this problem: the system plays or trains against a copy of itself. If it gets better, the opponent automatically gets better too — because opponent and trainee are the same model. This cycle can run without pause and without any human involvement.

Why self-play is so powerful

Human training has a natural ceiling: at some point, no human is capable of beating the model or correcting it in any meaningful way. Self-play knows no such limit. The system can, in theory, surpass human ability indefinitely, as long as the task has a clear success criterion — for example, a game rule that unambiguously states who has won.

On top of that comes sheer speed. A human might manage a few dozen training matches a day. An AI system plays millions of rounds against itself in the same amount of time. This allows for a breadth of experience that no human trainer could ever provide.

The mechanism behind the cycle

At the start there is a model that still plays randomly or very poorly. It competes against a copy of itself. After each match, it evaluates which moves or decisions led to victory, and adjusts its internal weights — that is, its stored experience values — accordingly. The copy is then updated to the new state, and the game starts over.

Importantly, the opponent is not always immediately brought up to the same level as the learning model. Older versions are often kept around to avoid excessive specialization. If two absolutely identical copies faced each other, they would only learn to beat the current version — not general strategies.

Technically, self-play is a form of reinforcement learning, a type of training in which the model learns through reward signals. What is special here is that this reward comes exclusively from the outcome of the match — no human has to say which move was good.

Self-play in products and headlines

The best-known application is AlphaGo, which in 2016 defeated the world’s best player at the board game Go. Its successor, AlphaZero, learned Go, chess, and shogi purely through self-play — without ever having seen a single human game. After just a few hours of training, it surpassed human world champions in all three games.

Self-play is no longer limited to board games. OpenAI used it to train robotic arms to manipulate cubes in a simulated environment. The principle also shows up in large language models: newer training approaches have models evaluate their own answers and weigh them against each other — a kind of self-play without a classic opponent.

One important limitation remains, however: self-play only works where a clear quality criterion exists. For a chess program, this is obvious — won or lost. For tasks like “write a good essay,” no such criterion exists. This is why self-play is harder to implement in the language domain and remains an active field of research.