AI Chess Bot Allie Learns Human-Like Play from 91 Million Lichess Games

Yiming Zhang, a Ph.D. student at Carnegie Mellon University’s Language Technologies Institute, created Allie, an AI chess bot that learns to play like a human by studying 91 million games from the online chess platform Lichess. Zhang, who began playing chess during the pandemic after watching "The Queen's Gambit," found existing chess bots frustrating because their moves often felt unnatural and incomprehensible to human players. This inspired him to build a system that emulates human thinking in chess. Allie differs from traditional chess engines, which are designed solely to win by simulating vast numbers of future moves through self-play. Instead, Allie was trained using a method similar to large language models like ChatGPT, but with game transcripts from real human players as its training data. This allowed Allie to learn not just strong moves, but also how humans approach the game—taking time to think in critical positions, making mistakes, and knowing when to resign. Zhang emphasized that the goal was not to create a superhuman player, but to build an AI that behaves in ways that feel authentic and relatable. “Before Allie, no chess engine modeled how people actually think,” he said. “They made instant moves in complex situations where humans would pause, or kept playing in hopeless positions—something most people wouldn’t do.” The team, including advisors Daphne Ippolito and Daniel Fried, combined classic AI search techniques with models of human behavior, resulting in a system that outperforms either approach alone. This hybrid method has potential beyond chess, with applications in therapy, education, and medicine, where AI agents that act in human-like, thoughtful ways could be more effective and trustworthy. Allie is open source and has already been used in nearly 10,000 games on Lichess. The project was presented at the 2025 International Conference on Learning Representations in Singapore, a top venue for machine learning research. Collaborators included Athul Paul Jacob from MIT and Vivian Lai from Visa. The researchers believe that studying how people interact with human-like AI is just as important as improving performance. “We’re not just building smarter machines—we’re building ones that understand and reflect human cognition,” said Ippolito. “That’s a powerful direction for the future of AI.”