The project, which received a best paper award, demonstrated that a certain bias in humans who train intelligent agents significantly reduced the effectiveness of the training.