This study examines voluntary corrective actions of advanced English learners teaching American students in science labs. Although the nonnative teaching assistants (TAs) possess sufficient knowledge in content areas, their level of mastery of the medium of the interaction hardly reaches that of their students. By describing their self-selected corrections on the microethnographic level, I aimed to reveal which aspects of language use the nonnative speakers have the most difficulty within the second language development. Analysis of the videotaped data of 5 Korean TAs in science labs of an American university revealed that their self-correction functioned to serve as critical means to gain control of L2 production. They tended to avoid uttering constructions with inanimate subjects and to depend on Korean topic-comment structure. Also, in many cases, Korean TAs’ troublesome production was repaired only after their student’s turn, so that the proper production was made across multiple turns. Finally, other semantic resources such as gesture contributed to the TAs’ communicating with their cognition, which evidences that talk in interaction is not separated from its relevant environment. These findings emphasize the importance of teaching advanced learners how to properly enact the assigned role by means of different resources including the language.