Virtual user model that operates machine spontaneously and makes errors is now developing as a novel usability test method. Conventional estimation methods of usability of artifacts have difficulties on practicability. Experiment with real-human subjects takes money and time. Also, most of conventional 'user model' do not simulate process of error emergence and recovery form error. Our Digital Human That Errs (DHTE) model makes error like humans and tries recovering, so that plausible user behavior can be obtained. Since characteristic of DHTE is controllable, weak points of machine can be detected with fewer simulations by reducing abilities of DHTE.