Near-random behavior in strategic games.

Abstract

Humans struggle to produce random behavior when asked to do so, leading some to suggest that our ability to randomize is attenuated (or even absent). Here, we demonstrate that the capacity for random behavior is much greater than previously acknowledged: Humans can behave in (near) optimally random ways when the context demands it. We created a custom, online version of the game Battleship in which participants competed against one another. On each turn, one player chose a tile on a 3 × 3 grid to place a ship, while the other player tried to guess the ship's location. Across six preregistered studies, this strategic context led to dramatic increases in random behavior relative to closely matched non-strategic tasks. Such random behavior was also remarkably flexible and optimal depending on context; both players readily incorporated varying information about the game (e.g., that tiles can differ in value) while simultaneously accounting for their opponents' knowledge of that information. A computational model validated these findings and revealed nuanced differences across contexts. Contrary to decades of work finding that humans fail to randomize, we find that strategic games promote highly flexible, near-random behavior.