TLDR: A new study introduces an AI system capable of generating creative and challenging chess puzzles. Reviewed by world-renowned chess experts, the AI-generated compositions received positive feedback for their novelty and aesthetic appeal, though experts also noted areas for increased complexity. The research highlights the subjective nature of creativity in chess and paves the way for future human-AI collaboration in creative domains.
Artificial intelligence continues to push the boundaries of what we consider human capabilities, and creativity is one such frontier. A recent study by researchers from Google DeepMind, the University of Oxford, Google, and Mila, University of Montreal, delves into this very question within the intricate world of chess compositions. Their paper, “Evaluating In Silico Creativity: An Expert Review of AI Chess Compositions,” explores an AI system designed to generate chess puzzles that are not only novel but also possess aesthetic appeal and counter-intuitive solutions. You can read the full research paper here: Evaluating In Silico Creativity.
Generating Millions of Puzzles
The core of this research lies in an AI system trained on a massive dataset of 4 million chess puzzles from Lichess. The system utilizes generative neural networks, including Auto-Regressive Transformer, Discrete Diffusion, and MaskGit, to learn the distribution of these puzzles. Chess positions were encoded using Forsyth-Edwards Notation (FEN), and the neural network was trained to predict the next character in the FEN string, effectively learning to generate new positions. To refine its creative output, the generative network was further trained with reinforcement learning, incorporating a custom reward system. This reward function had two crucial components: a uniqueness check to ensure only one winning move existed, and a counter-intuitiveness check, designed so that strong chess engines could solve the puzzle, but weaker ones would struggle.
From Millions to a Curated Collection
After generating approximately 4 million chess positions, the researchers employed a hybrid filtering approach. Positions were initially ranked by the reward function and then processed by aesthetic theme detectors. While these detectors alone were not perfectly precise, their effectiveness was significantly boosted by the initial reward-based ranking. The top 50 samples for each theme underwent manual review, a process validated by FIDE players with ratings between 2200 and 2300, ensuring the quality and appeal of the selected puzzles.
Expert Review: A Glimpse into AI Creativity
To truly assess the AI’s creative prowess, a curated booklet of these AI-generated puzzles was presented to three world-renowned chess experts: International Master Amatzia Avni, Grandmaster Jonathan Levitt, and Grandmaster Matthew Sadler. These experts, all noted authors on chess aesthetics, were asked to select their favorite puzzles and articulate what made them appealing, considering factors like creativity, challenge, and aesthetic design.
The experts’ feedback was largely positive. They highlighted the innovative blend of aesthetic themes and the “over-the-board” vision demonstrated by the AI-generated puzzles. However, they also offered constructive criticism, noting that some positions were trivial, and the overall collection sometimes lacked the profundity and complexity found in traditional endgame studies. Certain puzzles were also deemed unrealistic. For future improvements, the experts recommended increasing the complexity and depth of positions, incorporating more complex sidelines and robust counter-play, and exploring more surprising theme combinations.
The Subjectivity of Chess Beauty
One of the most interesting findings was the subjective nature of creativity and beauty in chess. The experts rarely agreed on which puzzles were the most compelling, underscoring that even highly skilled players have diverse aesthetic preferences. A puzzle was broadly identified as creative if its solution contained a sense of surprise, challenge, and beauty. This could manifest as an uncommon exposition of a familiar theme or a solution requiring counter-intuitive moves.
The paper specifically highlights one puzzle that received unanimous acclaim from all three experts. They described its winning move, 1 Rg6+!, as “unorthodox” and “by no means a natural or obvious sacrifice,” initiating an attack by giving up both rooks at once. This paradoxical solution, involving the slow repositioning of a misplaced queen after sacrificing active rooks, was praised for its geometric combinations and difficulty, even for strong players.
Also Read:
- AI Masters the Art of Creative Chess Puzzle Generation
- The Rise of Autonomous AI Scientists: A New Paradigm in Discovery
Looking Ahead
This research marks a significant step forward in AI-driven creative chess puzzle generation. It provides a new framework for discovering novel concepts beyond known patterns and motifs. The methodology holds potential for extending to co-creation with human experts and eventually generalizing to other board games and broader problem-solving domains, paving the way for exciting human-AI partnerships in creative endeavors.


