AI models have made significant strides in solving puzzles, yet they still falter on certain intelligence tests, according to technologyreview.com. While some models nearly mastered the New York Times Connections puzzles by early 2025, they continue to struggle with subtle variations in classic riddles and visual puzzles, highlighting ongoing limitations in AI cognition.
The development of AI puzzle-solving dates back to 1959 when IBM scientist Arthur Samuel introduced machine learning through a checkers-playing algorithm. Since then, games like chess and Go have served as benchmarks for AI progress. A Columbia University study in late 2024 found that leading AI models could solve only 18% of Connections puzzles, but by early 2025, improvements allowed some models to solve these puzzles almost perfectly. Despite this, certain puzzle types still expose AI weaknesses, especially those involving visual reasoning and nuanced changes.
These puzzle challenges reveal fundamental differences between human and machine cognition. While AI excels at pattern recognition and computation, it often fails where humans use intuition or contextual understanding. The mixed results underscore that AI, despite rapid advances, has not yet achieved human-like intelligence in all domains. Testing AI with puzzles remains a valuable method to assess its evolving capabilities and identify areas needing improvement.
Technologyreview.com offers readers the opportunity to try puzzles that have historically stumped AI models, illustrating where machines lag behind human intelligence. This hands-on approach serves as a practical demonstration of AI’s current strengths and weaknesses in problem-solving.