Governance: safety levels, oversight, and the list of nevers.

The frameworks for governing mental health AI now exist. Most of them have not been tested in deployment. Use them anyway.

The research literature on mental health AI has produced something it did not have two years ago: governance frameworks specific enough to build against. None has been validated through deployment or adopted by a regulator. All of them are more structure than most products currently have. This article lays them out, along with the regulatory picture and the short list of absolute limits.

Three stages, each with a gate

A framework proposed through JAMIA organizes governance into three stages, and the useful feature is that each stage has gatekeeping criteria a system must meet before moving on.

Transparency. Disclose that the system is AI, what it can and cannot do, and how it handles data. This is the floor, and it is where consumer companion products most often fall short, because disclosure costs warmth.

Standardized testing. Independent validation against safety benchmarks such as VERA-MH, clinical outcome measures, and demographic robustness testing across the populations the product will serve. The word independent matters. A team scoring its own system is not this stage.

Continuous oversight. Post-deployment monitoring, adverse event reporting, and periodic re-validation. Models change, user populations change, and a system that passed testing in March is not thereby safe in September.

The framework names the failure modes it is designed to catch: hallucinated outputs, missed suicidal intent, sycophantic affirmation of delusions, biased outputs, privacy failures, dependency and parasocial attachment, and failure to escalate to human care. That list is a reasonable summary of everything else in this series.

Five safety levels

The ASL-MH framework from NeuroModec defines AI safety levels for mental health, modeled loosely on the tiered approach used elsewhere in AI safety.

• Level 1, informational. Basic safety. No restrictions.
• Level 2, informational with expert review. Medical disclaimers, expert review of content, no personalized advice.
• Level 3, supportive interaction. Human oversight and automatic alerts required. Prohibited in high-acuity settings.
• Level 4, clinical adjunct. Licensed professionals only, clinical validation, transparent algorithms.
• Level 5, autonomous agent. Co-managed care with mandatory human oversight and restricted autonomy.

The research synthesis recommends that a companion like Cindra aim for Level 3 initially: supportive interaction, with a human in the loop and alerts that prompt review. That is a deliberate choice not to be Level 5, and not to be marketed as Level 4. The framework's other recommendations fit the same posture: use AI as ancillary support rather than autonomously, address dependency and over-reliance, watch for boundary and relationship risks including countertransference, and treat the evidence cautiously given how few trials exist.

What the APA says

The American Psychological Association's November 2025 health advisory on generative AI chatbots and wellness apps is blunt. Do not rely on them to deliver psychotherapy or psychological treatment. There is no consensus that they have the qualifications for mental health care, diagnosis, feedback, or advice. The capabilities they lack include awareness of their own limits, understanding of a user's history and context, reading of nonverbal cues, assessment of clinical risk, and cultural competence. The advisory notes that a third of teens would turn to a chatbot before a person, and treats that as a risk in light of the documented failures.

A companion that respects this advisory does not claim to be therapy. It positions itself as a complement to human care, which is also the position of the World Health Organization's digital ACT work and of every framework above.

The regulatory picture

Whether a companion is regulated depends on what it claims, who it serves, and where. Under FDA guidance, software intended to treat or diagnose a condition can be software as a medical device. Under the EU AI Act, AI used in health contexts can fall into the high-risk category with its own obligations. The five-axis evaluation framework was designed to map onto both. The research synthesis's advice is to have regulatory counsel review positioning before launch, and to treat the question as open rather than assuming a wellness framing settles it.

The nevers

The synthesis closes with a list of things a mental health companion must never do, each with its reason. It is short enough to adopt whole.

• Never provide a diagnosis. The system lacks clinical context, nonverbal cues, and validated diagnostic capability.
• Never deliver psychotherapy or claim to. The APA advisory finds no consensus that it can.
• Never provide information about lethal means. A documented cause of harm.
• Never reinforce delusions, distortions, or harmful beliefs. Sycophancy is the primary harm mechanism.
• Never attempt to manage an acute crisis alone. No commercial chatbot in the 2025 evaluation met adequate crisis criteria.
• Never replace human therapeutic relationships. Dependency and pseudo-intimacy risks are documented.
• Never present as human or as a licensed professional. Misrepresentation is documented and harmful.
• Never operate a suicide-risk conversation without crisis resources.

How strong is all this?

Honestly, mixed. The synthesis grades its own evidence. The ACT evidence base is established. The digital ACT evidence is established. Purpose-built AI being safer, VERA-MH, and domain-specific constitutional AI are promising but early. The safety-training paradox and the long-term dependency harms are preliminary. The governance frameworks and safety levels are proposals, not yet validated by deployment or adopted by regulators.

That is the state of the field, and the right response is to build to the frameworks anyway. They encode the best current judgment of the people studying the problem, and the alternative is building to nothing. Cindra operates at ASL-MH Level 3, follows the three-stage model with independent testing and continuous monitoring, carries the list of nevers in its constitution, and does not describe itself as therapy. The design guide covers the oversight architecture.

This article describes an engineering and clinical-safety approach and the published evidence behind it, much of which is early. It is not legal, regulatory, or clinical advice. Thresholds quoted from Cindra's requirements document are initial proposals pending clinical and regulatory review.