Playing in the Dark: Invisible Chess as a Laboratory for Strategic AI

ABSTRACT Artificial intelligence research has historically relied on games such as chess as benchmarks for progress in strategic reasoning. While chess and Go domains have produced remarkable advances—from Deep Blue to AlphaZero—they share a fundamental property that limits their relevance for many real‐world applications: perfect information. In contrast, most strategic environments in domains such as international relations are characterized by incomplete, uncertain, and asymmetric information. This perspective article makes three contributions. First, it proposes a structured taxonomy of imperfect‐information chess variants—including Kriegspiel and related “invisible board” environments—organized by the type and degree of information asymmetry they impose. Second, it presents an empirically grounded case study of Darkboard, the strongest computer Kriegspiel player to date, whose decade‐long record on the Internet Chess Club and three Computer Olympiad gold medals provides concrete evidence of both the power and the structural limits of AI under adversarial uncertainty. Third, it derives governance implications for the design and evaluation of AI systems operating in adversarial real‐world environments, where an opponent can model and exploit the inference procedure of an AI agent. We examine how the digital transformation of chess has shaped the development of AI methods and highlight the epistemic limitations of perfect‐information benchmarks. We discuss how games with imperfect information introduce challenges central to modern AI systems, including belief‐state reasoning, partial observability, and adversarial uncertainty. Finally, we outline implications for the future of AI innovation, suggesting that imperfect‐information strategic games may serve as valuable interdisciplinary laboratories for exploring the design, governance, and evaluation of intelligent systems operating under uncertainty.