Tuberculosis is a major public-health problem in South Africa, made worse by poor adherence to and frequent interruption of treatment. Direct observation (DO) of tuberculosis patients taking their drugs is supposed to improve treatment completion and outcome. We compared DO with self-supervision, in which patients on the same drug regimen are not observed taking their pills, to assess the effect of each on the success of tuberculosis treatment. We undertook an unblinded randomised controlled trial in two communities with large tuberculosis caseloads. The trial included 216 adults who started pulmonary tuberculosis treatment for the first time, or who had a second course of treatment (retreatment patients). No changes to existing treatment delivery were made other than randomisation. Analysis was by intention to treat. Individual patient data from the two communities were combined. Treatment for tuberculosis was more successful among self-supervised patients (60% of patients) than among those on DO (54% of patients, difference between groups 6% [90% CI -5.1 to 17.0]). Retreatment patients had significantly more successful treatment outcomes if self-supervised (74% of patients) than on DO (42% of patients, difference between groups 32% [11%-52%]). At high rates of treatment interruption, self-supervision achieved equivalent outcomes to clinic DO at lower cost. Self-supervision achieved better outcomes for retreatment patients. Supportive patient-carer relations, rather than the authoritarian surveillance implicit in DO, may improve treatment outcomes and help to control tuberculosis.