The acquisition of online interaction competencies is an important learning objective. The present study explored the relationships between the first-language heterogeneity of computer-supported collaborative learning (CSCL) groups and the development of students' online interaction self-efficacy via a pretest--posttest design in the context of a nine-week CSCL course. The research participants were 1525 freshmen receiving distance education who were randomly assigned to 343 CSCL groups. Independent of their own language status, students in CSCL groups featuring first-language heterogeneity exhibited lower precourse--postcourse gains in online interaction self-efficacy than students in groups without heterogeneity. Consistent with a theoretically derived moderation model, the relationships between first-language heterogeneity and self-efficacy gains were moderated by the amount of time that the groups spent on task-related communication during the initial collaboration phase (i.e., the relationships were significant when little time was spent on it but not when a great deal of time was spent on it). In contrast, the amount of time that groups spent on communication related to getting to know each other was ineffective as a significant moderator. Follow-up analyses indicated that time spent getting to know each other in first-language heterogeneous CSCL groups seems to have had the paradoxical effect of increasing rather than decreasing perceptions of heterogeneity among group members. Apparently, this effect impaired online interaction self-efficacy gains.