This mixed-methods study tested how LLM-assisted writing affects writers’ individuality. Fifty-nine students wrote on the same topic under deliberation (Sessions 1,3) and LLM assistance (Session 2). Surface convergence used Top-50 n-gram overlap and cumulative coverage curves; semantic cohesion used SBERT → PaCMAP and cosine-similarity heatmaps. Six retrospective interviews were coded in NVivo with Cohen’s κ. LLM assistance increased Top-50 overlap and steepened coverage curves; in embedding space, cohesion rose (Silhouette = 0.58) and the condition effect was significant (PERMANOVA p < .001). During the transition (LLM → deliberation), high-LLM users showed more frequent, longer blocks.