When ease of use is not enough: self-efficacy and continued use of AI educational tools in resource-constrained universities.

Generative artificial intelligence has entered second-language writing classrooms faster than applied linguistics has developed methods for measuring what such interaction changes in learner reasoning. ChatGPT can help learners generate ideas, reformulate sentences, organize drafts, and receive rapid language support, and recent research has reported promising uses in L2 writing, feedback, learner motivation, and classroom assessment (; ; ; ; ; ; ; ; ; ). These contributions are important, but they leave a more difficult question unresolved. If learners produce stronger essays after using ChatGPT, does this improvement reflect development in their own English reasoning, or does it partly reflect access to fluent external language? This question matters because writing quality and reasoning quality are not identical. A learner may produce a polished paragraph that is weakly justified, conceptually vague, or overly dependent on AI-generated phrasing. Conversely, a less fluent draft may reveal genuine development in claim formation, warranting, qualification, and counterargument. This distinction is especially important in EFL contexts, where learners may reasonably value ChatGPT for linguistic fluency while also becoming vulnerable to direct uptake, reduced authorial control, or unexamined acceptance of AI-generated claims (; ; ; , ). The central issue is therefore not whether ChatGPT can improve the surface form of learner writing. It is whether human-ChatGPT dialogue can support transferable growth in how learners reason in English. The present study addresses this issue through the domain of harm-sensitive artificial-intelligence literacy. This construct refers to learners’ ability to reason in English about AI-related ethical issues while distinguishing problem-solving intelligence from consciousness, suffering, moral standing, and entitlement. It is not treated here as a general attitude toward AI or as familiarity with digital tools. It is a language-mediated reasoning capacity visible in claims, warrants, hedges, contrasts, counterexamples, and explanations of harm. This focus is necessary because fluent language models can produce persuasive moral language without understanding, consciousness, or suffering in the human sense (; ; ). Learners who infer feeling, intention, empathy, or moral authority from fluent output may anthropomorphize AI systems in ways that weaken ethical reasoning. Anthropomorphism is therefore not only a philosophical concern. It is also a linguistic and pedagogical risk, since it appears in the verbs, predicates, and explanations learners use when writing about AI ().