This quasi-experimental study investigates the comparative efficacy of generative AI-enhanced self-learning and traditional Communicative Language Teaching (CLT) in Thailand’s low-proficiency English as a foreign language (EFL) context. The intensive exploration was conducted through pre/post-tests (TOEIC-modelled, α = 0.82) and a mixed-methods analysis of 36 students who shared similar English proficiency backgrounds. The findings reveal a statistically significant advantage for the teacher-led CLT approach. The CLT group (n = 16) attained a post-test mean of 63.6% (SD = 9.4), substantially outperforming the AI group (n = 20) at 38.7% (SD = 8.1), with a large effect size (Cohen’s d = 1.12). Qualitatively, while AI was valued for its flexibility (33.3%) and grammar support (16.7%), students eagerly reported critical shortcomings in developing oral proficiency and a sense of dehumanised interaction (19.4% cited loneliness). Conversely, CLT was strongly associated with essential social engagement (41.7%) and immediate feedback. Grounded in sociocultural and self-determination theories, these findings underscore AI’s current inability to provide human-mediated scaffolding in low-proficiency settings. The study concludes by advocating for a context-sensitive hybrid model integrating AI’s adaptive features with CLT’s interpersonal strengths, cautioning against AI-centric approaches in foundational language acquisition. This can suggest that AI is most effective when carefully scaffolded by human instructors who can compensate for its social and contextual limitations, and has important implications for teacher training programs, which may need to incorporate technology mediation skills alongside traditional pedagogical methods.