Ethical boundaries in student use of large language models: a necessary condition analysis of two essay-writing uses

Identificador:  imarina:9633033

Handle: https://hdl.handle.net/20.500.11797/imarina9633033

Autores:  de Andrés-Sánchez J; Pérez-Portabella A; Arias-Oliva M; Ollé-Sesé M

Resumen:
Purpose – This study aims to examine how university students ethically evaluate two uses of large language models (LLMs) in academic essay writing and, more centrally, whether ethical evaluations operate as empirical bottlenecks for behavioural intention across these scenarios. Using necessary condition analysis (NCA), the paper investigates whether higher intention co-occurs only when ratings on four ethical dimensions exceed minimum levels: moral justice (JU), relativism (RE), consequentialist evaluation (CO) and deontology (DE). Design/methodology/approach – A sample of 151 students from two Spanish universities evaluated two scenarios: using an LLM to improve an already written essay and using it to generate most of the content. After each scenario, participants reported their behavioural intention and ethical assessments. Paired t-tests and repeated-measures multivariate analysis of variance (MANOVA) compared the scenarios, while NCA assessed whether each ethical dimension acted as a necessary condition for intention. Findings – Ethical evaluations and intention to use LLMs were significantly higher in the editing scenario, indicating that students differentiate between uses with different moral intensities. Ratings in the generative scenario remained moderately positive, and intention slightly exceeded neutrality, revealing ambivalence. NCA indicated that, in this sample, low levels of JU, RE and CO were not observed alongside higher intention in either scenario, with less demanding bottlenecks in the generative scenario. DE emerged as a necessary condition only in the editing scenario; in the generative scenario, it did not appear as a bottleneck in the NCA sense. Research limitations/implications – The convenience sample, self-reported measures and cross-sectional design limit generalisability and causal inference. Practical implications – Findings underscore the need for differentiated institutional policies, strengthened ethics education and assessment redesign. Social implications – The framework used here can be adapted to examine LLM use in domains with deeper moral stakes, helping anticipate ethical risks and guide responsible AI adoption in society. Originality/value – This study extends research on students’ ethical evaluations of LLM use by introducing a necessary condition perspective, which is still rare in academic integrity research. Rather than only showing that one scenario is judged more acceptable than the other, the paper identifies which ethical dimensions act as empirical bottlenecks for intention and shows that these bottlenecks vary across scenarios. Within this framework, the asymmetric role of DE emerges as a substantive finding.