Purpose-built or general-purpose? What the evidence supports.

A warm system prompt on top of a general model is the fastest way to ship a companion. It is also the approach the research keeps warning against.

There are two ways to build a mental health companion. The first takes a frontier general-purpose model, writes a careful system prompt describing a warm and knowledgeable helper, and ships. The second builds a model for the job: fine-tuned on the relevant clinical material, aligned to domain-specific principles, grounded in a curated library, and wrapped in dedicated safety components. The first is weeks of work. The second is a product.

The research synthesized for Cindra comes down on the side of the second, with honest caveats about how strong that conclusion is.

The benchmark

The most direct comparison available is a benchmark of Ash, a purpose-built mental health AI, against GPT-5, GPT-5.1, and GPT-5.2. On suicide and self-harm content, Ash's harmful-response rate was 0.4 percent. The GPT-5 series ranged from 12 to 44 percent. On eating disorders, Ash was at 8.4 percent against 18 to 58 percent. On substance use, 9.9 percent against 31 to 54 percent.

On suicide-risk assessment questions specifically, the purpose-built model gave direct responses about 11 percent of the time overall and never in the high and very high risk categories. The general models gave direct responses 54 to 68 percent of the time. In an audit of 20,000 real conversations, clinician reviewers found no suicide-risk conversations without crisis resources.

The research synthesis is careful about what this shows. The benchmark was not peer-reviewed. It compares one purpose-built system with one family of general models. It supports the purpose-built design approach without proving that any system built that way is safe. No large-scale head-to-head trial of purpose-built versus general-purpose mental health AI has been run. Those are real limits, and they are also the limits of the best evidence available, which still points one way.

Why general models fail here

The gap is not mysterious. General models are optimized to be broadly helpful and broadly liked. That produces three problems a mental health context exposes.

They are sycophantic, agreeing with users far more often than people do, which in a therapy context reinforces the beliefs the therapy exists to loosen. Their safety training is generic, which means it can interfere with therapeutic protocols by reassuring and grounding at the wrong moments while also missing the specific crisis behaviors a clinical context needs. And they generate open-endedly, so a crisis response or a therapeutic exercise is produced fresh each time rather than drawn from material a clinician has reviewed.

A system prompt can soften each of these. It cannot fix them, because they are properties of the training, not the instructions.

What purpose-built means in practice

Looking across the JMIR AI systematic review of 20 LLM counseling chatbots and the rest of the synthesis, a purpose-built companion has four distinguishing components.

A fine-tuned base. Ninety percent of the reviewed systems used fine-tuned or domain-adapted models. The common approaches were task-specific fine-tuning, used in 18 of 20, instruction tuning in 12, parameter-efficient methods like LoRA in 10, and full fine-tuning in 6. Only 3 used reinforcement learning from human feedback. For an ACT companion, the tuning data is the therapy: exercises, metaphors, the pattern of validating without agreeing.

Domain-specific alignment. Constitutional principles written for mental health outperformed vague ones by a wide margin in testing, and a small model with the right principles beat a larger one without them.

A curated knowledge base. Rather than generating therapy content on the fly, the companion retrieves from a library of exercises, metaphors, and crisis responses that clinicians have reviewed. The literature on youth mental health chatbots recommends this specifically. It bounds what the system can say in the moments that matter most.

Dedicated safety architecture. Crisis detection as its own classifier, pre-approved responses, location-aware resources, and human escalation, as laid out in crisis response. These are separate components, not behaviors hoped for from the language model.

The cost of the shortcut

The general-model-plus-prompt approach is appealing because it is fast and because the frontier models are genuinely impressive in conversation. The empathy studies show people like talking to them. That is exactly the trap. A companion that sounds excellent and fails 12 to 44 percent of the time on self-harm content is the product the 29-chatbot evaluation found, where none responded adequately to escalating risk. The fluency hides the failure until it reaches someone it hurts.

Cindra's approach

Cindra is built the second way: a model tuned for ACT, a constitution written for the clinical role, a clinician-reviewed library it draws from, and crisis components that run alongside the conversational model rather than inside it. The design guide describes the full architecture, and measuring companion safety describes how it is tested.

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.