Auteurs : Hichour Hadjira . Chikhaoui Ahmed .
Students perform well with large language models (LLMs) and less well without them. This theoretical paper asks what learners internalize when the "more knowledgeable other" is an LLM. Drawing on Vygotskian and activity theory, scaffolding studies in educational technology and sociocultural second language research, it defines scaffold-dependent proficiency as a gap between performance with and without an LLM that does not close with practice yet is credited to the learner. It argues that default LLM use produces this state in three ways: its help works regardless of the learner's stage of development, it takes over the actions learners would otherwise perform while leaving them in control of when help arrives, and it separates the finished product from the learning the product was meant to show. Learners, the paper proposes, internalize their own share of the work and the pattern of the exchange, not the functions the LLM performs. The argument is developed through the case of English as a foreign language (EFL) writing at Algerian universities, where the linguistic operations LLMs perform most fluently are those students most need to learn. Implications for design, learning analytics and assessment follow.
Scaffolding ; Zone of proximal development ; Internalization ; Large language models ; Generative AI ; EFL writing ; Algeria ; Learning analytics