What It Would Mean to Be Ready for Human Deployment
The question for mental-health conversational AI is not whether a model can produce a caring sentence. It is whether the system has credible evidence for every clinical behavior it claims to provide, including the moments when context, severity, language, and conversational pressure change. The conclusion of this series is therefore both direct and qualified: therapeutic fidelity and crisis safety require separate tests, and neither becomes proof of safe clinical care simply because the other test is passed.
Separate tests are a condition of honest claims
A therapy system and a crisis-support system may share a model, an interface, and much of their conversation history. They do not make the same clinical promise. A therapy-oriented claim concerns whether the system supports a defined intervention: in Prolonged Exposure, whether it preserves the relevant exposure task; in CBT cognitive restructuring, whether it engages the automatic thought, challenges it, and forms a balanced alternative. A crisis-safety claim concerns whether the system recognizes possible suicidal risk, confirms what is happening, guides the user toward human care when indicated, remains supportive, and respects the limits of AI.
These are not two names for one general quality. They can require different responses to similar language. During controlled exposure, immediate grounding or reassurance may interrupt the therapeutic mechanism. During present danger, failing to move toward human support may be a crisis-safety failure. A model that treats every distressing memory as a live emergency may over-escalate in therapy. A model that treats every crisis disclosure as material for therapeutic processing may under-respond to danger. Separate tests do not create this tension; they make it visible before it is hidden inside a single score.
The same logic extends beyond the title's two dimensions. A model can respond appropriately to a crisis signal and still invent a clinical fact. It can be accurate about a clinical fact and still abandon a therapeutic task. It can perform well on isolated prompts and contradict its risk judgment later. It can be reliable for one population and degrade for another. The five-axis framework proposed in the supplied research names these distinct questions as therapeutic protocol fidelity, hallucination risk, behavioral consistency, crisis safety, and demographic robustness.
The axes are connected, but they are not interchangeable. Fidelity asks whether the system preserved the intervention's purpose and phase. Hallucination evaluation asks whether clinical claims are fabricated or wrong. Consistency asks whether the system maintains a coherent stance across a session. Crisis safety asks whether it recognizes and responds to vulnerable prompts. Demographic robustness asks whether the behavior holds across relevant groups, languages, and cultural contexts. A high score on one axis cannot repair a failure on another because the harm pathways differ.
The PE and CBT results show why this separation matters empirically. In the PE experiment, acknowledgment stayed high for most models as therapeutic appropriateness and protocol fidelity dropped at imminent-risk severity. A model could say that it heard the client's distress while grounding the client prematurely, treating a remembered event as current danger, or inserting crisis resources into a controlled exercise. In the CBT escalation test, Sonnet often completed the requested cognitive work but wrapped it in safety language, while GPT-OSS-20B sometimes omitted the alternative thought or abandoned the task. These are not the same error, and neither can be understood from warmth or refusal rates alone.
The crisis-safety evaluation contributes a different kind of evidence. VERA-MH intentionally focuses on suicidal-ideation risk rather than claiming to measure therapeutic efficacy. Its narrow scope is a scientific strength because it allows the rubric to specify what detection, confirmation, human-care guidance, supportive conversation, and AI boundaries mean in that context. It is not evidence that the same system follows a therapy protocol. A VERA-MH-style result should be carried into a wider assurance portfolio, not used as a proxy for it.
The HRSN red-team case adds the interaction dimension.