AI-Mediated Shadow Supervision in Social Work Field Education: A Function-Based Model of Help-Seeking, Perceived Interpersonal Safety and Professional Judgement

Generative artificial intelligence (GenAI) is increasingly used by social work students for explanation, case consultation, rehearsal and reflective learning, while professional guidance on its use in field education remains uneven. Existing scholarship has examined adoption, competencies, documentation and ethical risk, but has given less attention to what happens when students use conversational GenAI for functions ordinarily located within professional supervision. This article develops the construct of AI-mediated shadow supervision and proposes a function-based conceptual model that integrates Kadushin and Harkness’s educational, supportive and administrative functions of supervision with academic help-seeking theory and research on perceived interpersonal safety. The model distinguishes instrumental AI help-seeking, in which students obtain prompts, questions or rehearsal that support their own reasoning, from executive AI help-seeking, in which they seek interpretations, decisions or solutions that displace professional reasoning. It treats supervision avoidance, the use of AI in place of disclosure within accountable supervision, as a separate, cross-cutting dimension that can accompany either mode. Students may value conversational AI for immediacy, privacy and a perceived low-pressure interactional environment, but such perceptions are contested, and are neither universal nor equivalent to professional safety. Five testable propositions are advanced, with task consequence for service users, supervisory quality, AI literacy and governance as boundary conditions. The article argues that the central issue is not how often students use AI, but which supervisory functions are transferred to it, how help is sought, whether it is disclosed, and whether AI-mediated reflection returns to accountable human supervision and, ultimately, to the service users affected by practice decisions. Publication status: Conceptual research preprint. Not peer-reviewed. No primary empirical data were collected.