Abstract  CAPTCHA has been widely deployed by commercial web sites as a security technology for purposes  such as anti-spam. A common approach to evaluating the robustness of CAPTCHA is the use of machine learning techniques. Critical to this approach  is the acquisition of an adequate set of labeled samples, on which the learning techniques are trained.  However, such a sample labeling task is difficult  for computers, since the strength of CAPTCHAs  stems exactly from the difficulty computers have in  recognizing either distorted texts or image contents. Therefore, until now, researchers have to  manually label their samples, which is tedious and  expensive. In this paper, we present Magic Bullet, a  computer game that for the first time turns such  sample labeling into a fun experience, and that  achieves a labeling accuracy of as high as 98% for  free. The game leverages human computation to  address a task that cannot be easily automated, and  it effectively streamlines the evaluation of  CAPTCHAs. The game can also be used for other  constructive purposes such as 1) developing better  machine learning algorithms for handwriting recognition, and 2) training people’s typing skills