Allie in AI Chess Bot learns to play like a human from 91 million lichen games

Yiming Zhang didn't grow by playing chess. Like many others, Carnegie Mellon University PhD students discovered the Netflix series “The Queen's Gambit” during the pandemic and began playing online. However, he quickly realized how unnatural it felt to play against the Chesbot.

“After learning the rules, I was in the bottom 10% and 20% of the players online,” said Zhang, part of the Language Technologies Institute (LTI), CMU's computer science school. “For beginners, playing against chesbots is neither interesting nor beneficial because the movements they make are often strange and incomprehensible to humans.”

Zhang's complaints have led him to develop Allie, an artificial intelligence-powered chesbot that demonstrates the benefits of AI tools that he thinks like a human. He believes that training future AI systems to reflect and deliberate complex issues could create better agents for use in therapy, education and medicine.

“I'm obsessed with building superhuman AI, which has better mathematics and other reasoning tasks than most people,” said Daphne Ippolito, Zhang's advisor and assistant professor at LTI. “But there are many opportunities to train AI models to act like humans. I think it's worth exploring.”

Ally plays like a human and can adapt to a variety of strengths, from beginners to experts. It was trained in the same way as the language models that support modern chatbots, such as ChatGpt. But instead of giving Allie Text from the Internet, the team trained it with 91 million transcripts from Lichess, a popular Chess platform. Exposing Allie to a human-played chess game transcript taught her how to create movements that human players make, spending time thinking about important positions and resigning when the game can't win.

“We are excited how the adaptation methods we used combine classic AI search procedures with modelling of human behavior, and how that combination is better than itself in either way,” said Daniel Fried, assistant professor at LTI who worked on the project. “The methods like the ones we used have already been applied to complex games such as diplomacy. We look forward to being used in other tasks that require AI to act strategically, but in a way that is used in human-compatible ways.”

Most chess engines are built with one goal to win. They simulate countless future movements, embracing each other's changes in a self-improvement loop with no human data. This approach brings systems with almost invincible strength, such as Alpha Zero and Stockfish, making them non-hospitable opponents for casual players and beginners.

“Before Ally, there was no chess engine modeling how people thought,” Zhang said. “The Chessbots would either move instantly in complex positions where humans needed time to consider different options, or continue to play in completely lost positions where humans normally resign. This made the chess AI that existed look unnatural.”

When asked about plans for the future, the team explained that Allie is completely open source, and that since its deployment on Lichess, it has accumulated nearly 10,000 games.

“Our project makes sense because it evaluates how people interact with AI that tries to become human,” Ippolito said. “We also intentionally built an open source platform that people can build.”

Ally was presented at the 2025 International Conference on Learning Expression in Singapore, one of the finest venues for machine learning research. Zhang, Ippolito and Fried collaborated with PhD Athul Paul Jacob on a project. A student at Massachusetts Institute of Technology. Visa researcher Vivian Rye.