Maia Chess is an open-source, human-like chess engine: a neural network trained on millions of human games that plays the moves a person of a given skill level would actually choose. Developed by the Computational Social Science Lab, the project has about 1,300 stars on GitHub and remains active in 2026.
Conventional chess engines try to play perfectly, which makes their moves alien to human players. Maia takes the opposite approach: trained on millions of online games, it learns to predict human decisions, including characteristic human mistakes. Instead of one model, Maia is a family of models, each tuned to a specific rating band from beginner to strong club player, and each predicts the move a player of that strength would make with accuracy far above chance. Built on the Leela Chess Zero framework, Maia has become a reference point in research on human-AI alignment and human-centered machine learning, and its bots let anyone play against a convincingly human opponent online.
Maia’s training data comes from millions of games played by humans online, and the Maia bots deployed on Lichess have played large numbers of games against the public, which makes it easy to experience the difference: the engine blunders, attacks and simplifies in recognizably human ways. Researchers have also used Maia as a tool for studying teaching and skill acquisition, because a model of a 1500-rated player is far more instructive to a 1500-rated student than a perfect engine.
Maia Chess is free and open source under the GPL-3.0 license. The code, models and research artifacts are available on GitHub at no cost, and the engines can be played online for free.
Maia Chess is for chess players who want a human-feeling sparring partner at their level, and for researchers studying how AI systems can model and align with human decision-making.