Ethical Justification and Selection Criteria for the Use of Local Large Language Models in Novice Counselor Training

This conceptual study proposes an evaluative framework for using locally deployed large language models (LLMs) in novice counsellor practice training, addressing two questions: whether cloud-based LLM use is ethically defensible in counselling training, and what criteria should govern local LLMs as an alternative.The study combines a narrative review with a written consultation of three field practitioners, from trainee to supervisor.The study argues that, for individual counselors and trainees without institutional legal resources, cloud-based LLM use is difficult to justify: protected data comes from the counselor's own records, possible admixture of real case fragments, and the tool's likely extension to transcript-based supervision, under Korean restrictions on cross-border data transfer.Four selection criteria are presented, with their sources differentiated-data locality and Korean-language performance drawn from ethics and development theory, and hardware suitability and license openness imposed as practical constraints-along with one operational design principle: theoretical consistency.Drawing on the deliberate practice paradigm, the study proposes pairing simulation with structured feedback as a key design requirement, since unassisted practice may degrade empathy.A review of Korean-capable models released through the first half of 2026, together with a preliminary on-device measurement of a representative model, confirms that a 7-8B-scale model actually runs on a consumer-grade GPU while showing limited accuracy in detecting subtle counselling errors; three safeguards are proposed to mitigate small-model limitations, together with the empirical agenda for a subsequent implementation study.