The debate over 24/7 AI access in mental health has been framed as a question of dependency risk — the concern that constant availability produces reliance, erodes resilience, and substitutes for genuine therapeutic work. This concern is clinically grounded and correct within its own ontology: human-to-human therapeutic relationships carry attachment dynamics, reciprocity, and emotional reinforcement that can produce dependency when made continuously available. But the debate has failed to distinguish between two categorically different interventions: therapeutic availability (relational, emotional, attuned, dyadic, identity-shaping) and operational availability (informational, procedural, stabilizing, bounded, non-relational). AI health agents that are designed and deployed as operational surfaces — not as relational actors — do not create dependency, because dependency requires reciprocity, attunement, and emotional reinforcement that a correctly designed agent does not provide. The risk is not 24/7 access. The risk is misclassification: products that frame AI as a relational actor, use therapeutic language, and market themselves as emotional support create dependency through design, not through availability. This paper argues that the correct distinction is between the two availability types, defines the properties of each, identifies the conditions under which operational availability is appropriate and non-harmful, and names the misclassification failure that makes 24/7 AI in mental health dangerous when it is dangerous.
Narnaiezzsshaa Truong (Tue,) studied this question.