Yiming Zhang, a PhD student at Carnegie Mellon University, used to play chess online but felt frustrated by chess bots (computer programs designed to play chess). These bots often made strange moves that beginners could not understand, making the game less fun and less helpful for learning.
Zhang created Allie, a new kind of chess bot using artificial intelligence (AI). Unlike typical bots that aim only to win by calculating millions of moves ahead, Allie learns from real human games. It was trained on 91 million game records from Lichess, an online chess platform. This training helps Allie make moves like a person would, think longer on hard positions, and give up when the game is clearly lost.
Allie can adjust to play at different skill levels, from new players to experts. This makes it a better opponent for learning because it feels natural. While many AI efforts focus on superhuman abilities, like solving math problems better than people, there is value in AI that acts like humans. This could help in areas such as therapy, education, and medicine, where AI needs to handle complex issues thoughtfully.
How Allie was built
Allie combines old AI search methods, which explore possible future outcomes, with models of human behavior. This mix works better than using just one approach. Similar ideas have been used in games like Diplomacy, where strategy must feel human.
Before Allie, chess engines like AlphaZero or Stockfish focused on unbeatable strength without human data, leading to unnatural play. They move instantly in tough spots or keep going in hopeless games, unlike people.
Allie is open source, meaning its code is free for anyone to use and improve. Since launching on Lichess, it has played nearly 10,000 games. The project was shown at the 2025 International Conference on Learning Representations in Singapore. The creators see Allie as a way to study how people interact with human-like AI and to provide a base for others to build on.