Trying to be unpredictable even against unobserved opponents shifts humans away from randomness but not from human predictability

In competitive settings, randomization can prevent exploitation by opponents, yet human choices depend on both opponents' past actions and players' own action histories. Why this latter dependence arises, and how it affects predictability to others, remain unclear. Across three experiments, we separated sequence generation from subsequent prediction, thereby preventing generators from adapting to predictors' responses. Whereas sequences generated under instructions to be random were biased toward alternation, those generated to evade prediction were biased toward repetition. Nevertheless, these sequences remained predictable. Human predictors systematically over-predicted repetitions, so sequences with a slight bias toward alternation actually minimized predictability. Thus, the failure to evade prediction may arise not merely from deviations from randomness, but also from inaccurate estimates of how others would predict one's behavior. These findings suggest that understanding human competitive behavior requires considering not only how people adapt to others' actions, but also the human-specific biases emerging in these processes.