AI Adapts Less Readily to Video Games Than Board Games, Underscoring General-Learning Challenge

Board games have often served as milestones of AI capability since IBM's Deep Blue defeated world chess champion Garry Kasparov in 1997. New York University's Game Innovation Lab at the Tandon School of Engineering says board games have closed rule sets and can be simulated quickly. Modern video games also test spatial reasoning, long-term planning, trial and error, and social intuition, making them closer to the variability of the real world.

NYU published the research perspective from Julian Togelius and colleagues on March 25, 2026, and related coverage appeared in Taiwan on April 2. The team said reinforcement learning often relies on millions or even billions of simulations, yet performance can collapse after minor changes to rules, screen colors or object positions. A genuinely general system should be able to learn an unfamiliar game from scratch within dozens of hours, much like an experienced human player.