Randomized controlled trials have proliferated in education, in part because they provide an unbiased estimator for the causal impact of interventions. It is increasingly recognized that many such trials in education have low power to detect an effect if indeed there is one. However, it is less well known that low powered trials tend to systematically exaggerate effect sizes among the subset of interventions that show promising results (p < [alpha]) We conduct a retrospective design analysis to quantify this bias across 22 such promising trials, finding that the estimated effect sizes are exaggerated by an average of 52% or more. Promising trial bias can be reduced ex-ante by increasing the power of the trials that are commissioned and guarded against ex-post by including estimates of the exaggeration ratio when reporting trial findings. Our results also suggest that challenges around implementation fidelity are not the only reason that apparently successful interventions often fail to subsequently scale up. Instead, the effect from the initial promising trial may simply be exaggerated.