This qualitative study examines how bilingual EFL students used their first language (L1) when interacting with ChatGPT for academic writing and identifies forms of potential scaffolding generated by AI across the writing process. The participants were 40 undergraduate students enrolled in an academic writing course at a Vietnamese university. The dataset comprised 40 student - ChatGPT chat transcripts and 40 reflection journals produced as students completed an academic essay task. The data were analyzed using thematic and content analysis, supplemented by descriptive frequency counts. The findings show that 77.5% of the students switched between Vietnamese and English during their interactions with ChatGPT. L1 use was most prominent for idea generation during the prewriting stage (72.5%), followed by vocabulary translation (62.5%), clarification of sentence meaning (55%), and grammar checking during the postwriting stage (47.5%). ChatGPT provided five recurring forms of support: guiding questions (65%), bilingual vocabulary support (62.5%), sentence-structuring assistance (55%), translanguaging facilitation (50%), and grammar and syntax feedback (47.5%). The results suggest that students generally used Vietnamese as a cognitive and metacognitive resource rather than as a substitute for writing in English. However, 32.5% of the students reported that some guiding questions were overly generic, while others experienced excessively formal vocabulary suggestions and overly technical grammar explanations. The study highlights the value of stage-sensitive and critically mediated AI support in bilingual EFL writing instruction.