The development of effective interventions for COVID-19 vaccination has proven challenging given the unique and evolving determinants of that behavior. A tailored intervention to drive vaccination uptake through machine learning-enabled personalization of behavior change messages unexpectedly yielded a high volume of real-time short message service (SMS) feedback from recipients. A qualitative analysis of those replies contributes to a better understanding of the barriers to COVID-19 vaccination and demographic variations in determinants, supporting design improvements for vaccination interventions. Objective: The purpose of this study was to examine unsolicited replies to a text message intervention for COVID-19 vaccination to understand the types of barriers experienced and any relationships between recipient demographics, intervention content, and reply type. Method: We categorized SMS replies into 22 overall themes. Interrater agreement was very good (all κpooled . 0.62). Chi-square analyses were used to understand demographic variations in reply types and which messaging types were most related to reply types. Results: In total, 10,948 people receiving intervention text messages sent 17,090 replies. Most frequent reply types were “already vaccinated” (31.1%), attempts to unsubscribe (25.4%), and “will not get vaccinated” (12.7%). Within “already vaccinated” and “will not get vaccinated” replies, significant differences were observed in the demographics of those replying against expected base rates, all p . .001. Of those stating they would not vaccinate, 34% of the replies involved mis-/disinformation, suggesting that a determinant of vaccination involves nonvalidated COVID-19 beliefs. Conclusions: Insights from unsolicited replies can enhance our ability to identify appropriate intervention techniques to influence COVID-19 vaccination behaviors.