JMIR Publications

ABSTRACT

Background:

Conversational artificial intelligence (AI) is now an ordinary feature of adolescent help seeking, and users reportedly form alliance-like bonds with such systems. Debate about AI in mental health care has been dominated by capability benchmarking, and has been conducted almost entirely with reference to well-resourced Global North settings. Little is known about what these systems would have to substitute for in low-resource school settings, or in work with sexual minority learners, for whom disclosure to an dules in a school carries distinctive risks.

Objective:

This study examined how school-based counselling practitioners in South Africa describe their work with gay and same-sex attracted learners at risk of suicide, in order to specify the constituent properties of relational support, determine which are plausibly substitutable by current or foreseeable AI systems , and established where AI may and may not be legitimately deployed.

Methods:

We conducted a qualitative descriptive study using semi-structured individual interviews with 11 school-based counselling practitioners working in three South African provinces, purposively sampled on current practice and experience supporting learners who had disclosed same-sex attraction and suicidal ideation. Data were analysed using reflexive thematic analysis; reporting follows COREQ-32. Participants were not asked about artificial intelligence and none raised it, the comparison between their accounts and the AI capabilities is an analytic overlay contributed by the authors and declared throughout.

Results:

Five constituents of relational support were identified: trust as an accumulated and accountable achievement rather than a procedural formality; the reading of distress in embodied and paralinguistic cues; identify affirmations as a staged developmental process rather than a single affirming response; advocacy conducted beyond the counselling room within school and family systems; and the moral and emotional cost borne by the practitioner. A sixth, contextual finding cut across the others: participants described schools in which teaching and support staff were themselves sources of stigma and bullying, and in which no counselling resource existed at all. Mapping these against the AI literature yielded a three-part taxonomy of limitation: contingent-technical, which current systems are already overcoming; institutional-structural, durable but contingent on law and regulation; and constitutive, concerning reciprocal vulnerability and moral answerability, which do not reduce to technical capability.

Conclusions:

Arguments for the irreplaceability of human practitioners frequently rest on contingent-technical claims that are already obsolete, weakening an otherwise sound position. The durable case rests on institutional standing and on constitutive reciprocity. We propose a permissibility framework specifying where AI may operate independently, where it may operate only as an adjunct, and where it should not operate at all in school-based support for sexual minority learners, together with the safeguards each tier requires. We further argue that the common recommendation of AI as a bridge to human care presupposes that the human destination is safe, an assumption our data do not support in all settings. South Africa currently has no regulatory instrument governing AI-delivered psychological services; we set out what such an instrument would need to address