Yiming Zhang didn’t grow up playing chess. Like many others, the Carnegie Mellon Ph.D. student got interested after watching Netflix’s The Queen’s Gambit during the pandemic. He started playing online but quickly grew frustrated with existing chess bots.
“After learning the rules, I was in the bottom 10–20% of online players,” Zhang said. “For beginners, playing against bots isn’t fun or helpful. Their moves often seem strange and hard to understand.”
This frustration led Zhang to create Allie, a chess bot designed to play more like a human. Allie shows how AI can be built to think in human-like ways, which Zhang believes could also benefit fields like therapy, education, and medicine.
“There’s been a big push to make superhuman AI—systems better than people at math or reasoning,” said Daphne Ippolito, Zhang’s adviser and CMU assistant professor. “But it’s just as valuable to build AI that acts more like people, and that’s worth exploring.”
Unlike most chess engines, which are programmed only to win, Allie can adapt to different skill levels, from beginner to expert. It was trained like a language model but on chess data instead of internet text—specifically, 91 million transcripts from games played on Lichess. By learning from human games, Allie developed the ability to:
-
make moves that resemble human decision-making,
-
pause to think during tough positions, and
-
resign when the game is clearly lost.
“I like how our methods combine classic AI search with modeling human behavior,” said Daniel Fried, another CMU professor on the project. “This combination is stronger than either method alone. It has worked in games like Diplomacy, and it could also apply to other areas where AI needs to act strategically but in human-like ways.”
By contrast, most chess engines like AlphaZero or Stockfish become incredibly strong by simulating endless possible moves against themselves, without human data. While nearly unbeatable, they aren’t much fun for casual players.
“Before Allie, no chess engine really modeled human thought,” Zhang explained. “Bots moved instantly in complicated positions, when a human would take time. They kept playing hopeless games instead of resigning. That made them feel unnatural.”
Allie, now open source, has already been used in nearly 10,000 games on Lichess.
“Our project matters because it looks at how people interact with AI designed to feel humanlike,” Ippolito said. “And we made it open source so others can build on it.”
Press release from Carnegie Mellon University
