Parallelization of fitness evaluation is an established practice in evolutionary computation, and is a necessity in applications where fitness functions are computationally expensive. Traditional master-slave EAs based on a synchronous, generational model incur idle time when there is variance in the time it takes for individuals to have their fitness evaluated. Asynchronous evolutionary algorithms based on a steadystate model can make more efficient use of parallelization by eliminating idle time and reclaiming CPU resources. It is believed, however, that asynchronous EAs are biased toward regions of the search space where solutions take less time to evaluate, and away from regions where fitnesses evaluation is expensive. We show experimentally that asynchronous EAs do indeed exhibit an evaluation-time bias. This bias can either cause or prevent premature convergence. We also show, however, that on a flat fitness landscape, the asynchronous EA is attracted to both fast and slow regions of the search space, and away from medium-speed solutions. This indicates that further work is needed to understand the implications that asynchrony has for EA applications.