We Used It Anyway: Low Trust and High Adoption of Large Language Models in Requirements Engineering Education

Recent advances in Large Language Models (LLMs) are reshaping how students approach Software Engineering activities, particularly those centered on textual artifacts. This paper reports on the experience of integrating a non-restrictive LLM use within a first-year Master’s course in Software Engineering, in which students completed a semester-long Software Engineering project divided in a Requirements Engineering (RE) phase and a software testing phase. Following the completion of the RE phase of the project, we administered a 12-question survey to 43 students to investigate which LLM tools they used, for which activities, how they evaluated AI-generated outputs, and how they envisioned the role of LLMs in future Software Engineering curricula. Our findings show that LLM use was nearly universal among respondents and concentrated primarily on text-intensive activities such as use-case narratives, consistency checking, brainstorming, and refinement of existing artifacts. Despite this widespread adoption, most students perceived AI-generated outputs as inferior to their own work, frequently reporting issues related to incorrect or incomplete content, mismatches with course-specific conventions, over-engineering, and limitations in diagram generation. Nevertheless, students continued to rely extensively on LLM as a reviewer, critic, and productivity aid rather than as a fully autonomous artifact generator. We also observe that students’ choice of models was strongly influenced by access to premium subscriptions available for free for student accounts, highlighting the role of economic accessibility in shaping educational AI ecosystems. The results suggest that the most significant challenge in the adoption of LLMs within Requirements Engineering education is not whether students use these tools, but whether they possess the skills required to critically evaluate, contextualize, and correct its outputs.