This study examines how future teachers’ AI literacy develops throughout a four-module professional development program and how metacognitive reflection supports transformative learning. Sixty-three Italian teacher-learners produced short reflective diaries after each module. Using a convergent mixed-methods design, we integrated computational text analyses with qualitative coding to characterize change across cognitive, operational, critical, and ethical dimensions of AI literacy. Texts were pre-processed in Italian with spaCy (lemmatization and extended stopword filtering) and represented with TF-IDF to compute inter-section cosine similarity between phases and within-participant coherence; lexical diversity (type-token ratio), unsupervised topic modeling (LDA, k = 5), and sentiment scores complemented the analysis. Qualitative thematic analysis contextualized patterns with coder agreement and representative excerpts. Results indicate declining inter-diary similarity across modules alongside stable or rising within-participant coherence, suggesting cohort-level conceptual reorganization while individuals consolidate their understanding. Lexical diversity decreases modestly as vocabulary specializes; topic distributions shift from generic to practice-oriented themes; and sentiment evolves from uncertainty to more balanced, professionally grounded appraisals. Taken together, the findings portray a progression from initial ambiguity toward increasingly articulated AI literacy, linking computational signals with reflective evidence of transformation. We discuss implications for teacher-education curricula that aim to build sustainable, ethically aware AI capacities without presupposing technical backgrounds.