Crafting effective academic titles is a challenging task that requires balancing informativeness, conciseness, and reader engagement. This paper investigates titles produced by master’s students of Linguistics at the Faculty of Letters and Humanities of Sfax (FLSHS) and compares them with research article titles written by expert authors and with AI-generated alternatives. The study aims to evaluate students’ titles in relation to expert norms and to explore the potential of Large Language Models (LLMs) in academic title generation. A corpus of 659 titles, including master’s dissertations, research articles, and AI-generated titles, was analysed quantitatively and qualitatively using a synthesised model based on Ken Hyland and Zou (2022) and Swales and Feak (2012). The findings show that students adhere more closely to academic title conventions by producing informative and lexically dense titles, whereas expert authors prioritise reader engagement. AI-generated titles, although capable of producing useful lexical content, rely heavily on formulaic expressions and often fail to recognise the genre-specific conventions of academic discourse. They tend to be longer, more clausal, and more question-based than human-crafted titles. The study suggests that AI tools can serve as valuable brainstorming resources in academic writing pedagogy when their output is critically evaluated and adapted by users.