Readiness for AI-First Disclosure In Social Work: Competence Gaps, Predictors of Response Confidence, and Latent Profiles In Hong Kong

Background and Purpose

As conversational AI becomes a primary emotional outlet for distressed individuals, social workers are increasingly encountering clients who have disclosed to an AI system before or instead of seeking human help. This AI-first disclosure pattern raises distinct clinical demands: practitioners must assess what the client shared, how the AI responded, what relational meaning the client has attached to that interaction, and how to navigate a therapeutic alliance that now includes an AI as a prior confidant. Yet social work has not examined whether practitioners are prepared for this reality. Current workforce AI debates focus predominantly on professional tool adoption and digital literacy - competencies oriented toward using AI, not responding to clients who use AI. This study addresses that gap by examining: (1) whether practitioners feel equipped to respond to AI-first disclosure; (2) which competencies and contextual factors actually predict that confidence; and (3) what latent readiness profiles characterise the workforce.

Methods

A cross-sectional online survey was completed by 372 social workers in Hong Kong. Measures included self-rated competence and perceived importance across theoretical, practical, and ethical AI literacy domains; AI wariness; daily professional AI use; organisational disclosure protocols; training exposure; and direct experience with client AI-first disclosure. Analyses included descriptive comparisons, multivariable linear regression, and latent profile analysis.

Results

Over half of respondents (57.1%) had encountered client AI-first disclosure, yet only 34.5% felt confident in their ethical and clinical response, a preparedness gap that forms the study's central finding. Multivariable regression tested whether AI literacy predicted this confidence. It did not support the linkage: theoretical AI knowledge, general AI proficiency, and daily AI use were all non-significant predictors. What significantly predicted response confidence were ethical AI competence (β = .28, p < .001) and the presence of clear organisational protocols (β = .19, p = .002), pointing to practice-grounded ethics and institutional governance, not technical upskilling, as the critical readiness factors.

Latent profile analysis identified three practitioner subgroups (entropy = .882): an underprepared majority (~45%), a developing middle group (~42%), and a high-competence minority (~13%). A key dissonance emerged: the highest-readiness subgroup reported the least formal AI training, while the most trained subgroup showed only moderate competence, suggesting that existing training programmes are not producing clinical readiness for this specific challenge.

Conclusions and Implications

The assumption that AI literacy prepares social workers to respond to client AI use is not supported by these data. Confidence in navigating AI-first disclosure is shaped by ethical reasoning capacity and organisational readiness, not by knowing how AI works or using it professionally. For social work education and NGO training infrastructure, this reframes the priority: rather than expanding AI tool training, the field urgently needs ethically grounded, scenario-based preparation for AI-mediated help-seeking, alongside institutional protocols that give practitioners a clear clinical and procedural footing when a client's first confidant has been a machine.