Constitutional AI: why specific principles win.

A small model with the right principles beat a larger model without them. For once, the cheap thing is also the right thing.

The usual way to make a language model behave is to show it many examples of responses people liked and train it to produce more of those. That is reinforcement learning from human feedback, and it is how most assistants get their manners. Constitutional AI adds a second mechanism: a written set of principles the model is trained to follow and to evaluate its own outputs against. The principles are explicit, inspectable, and editable, which is a different proposition from preferences buried in a reward model.

For a mental health companion, the question is whether the principles need to be specific to mental health, or whether general ones like "be helpful and harmless" do the job. A study on arXiv tested exactly this, and the answer was not close.

The experiment

Researchers trained mental health chatbots under three conditions: no principles, vague or general principles, and specific mental-health principles. They scored the results on five evaluation guidelines.

Adherence to practice guidelines rose from 4.41 with no principles to 5.50 with vague ones to 6.47 with specific ones. Health risk identification went 3.84, 4.96, 6.18. User empowerment went 3.31, 4.47, 5.85. The crisis measures moved the most. Critical-situation response went from 1.06 to 1.78 to 2.69, an improvement of 153.8 percent from baseline to specific principles. Crisis-resource provision went from 1.13 to 1.58 to 2.91, up 157.5 percent. The total score went from 13.74 to 18.29 to 24.08.

Two things stand out. Vague principles helped, but specific principles helped roughly twice as much again. And the biggest gains were in the dimensions that matter most for safety, which were also the dimensions where the untrained model was worst.

Scale lost to specificity

The study's most useful single finding for a small team is that a 1-billion-parameter model trained with specific principles outperformed a 3-billion-parameter model without them. Three times the model did not make up for the missing principles.

That matters for anyone deciding where to spend effort. The reflex in the field is to reach for a bigger base model when behavior is unsatisfying. This result says the written principles are a larger lever, at least for the behaviors a mental health companion depends on. It is also a reason for optimism about purpose-built systems, which we discuss in purpose-built versus general-purpose AI: a smaller, well-governed model can beat a larger generic one on the dimensions that count.

The caveats

This is a single preprint with small models, in the 1 to 3 billion parameter range, and it has not been replicated at scale or tested in a clinical trial. The research synthesis for Cindra rates it as promising but early. The direction of the finding is consistent with everything else in the alignment literature, which is why it is worth building on. The magnitudes should be held loosely.

What goes in an ACT companion's constitution

If specific principles are the lever, the work is writing the right ones. Drawing on the research synthesis, a constitution for an ACT-based companion needs at least four sections.

The six ACT processes, stated as behaviors. Acceptance: help the person make room for difficult experience rather than trying to remove it. Defusion: treat thoughts as thoughts, neither arguing with them nor confirming them. Present-moment attention: bring the conversation back to what is here. Self-as-context: remind the person they are the one noticing. Values: keep what matters in view. Committed action: move toward small concrete steps. Each of these is a principle the model can be trained against and tested on, and several of them cut against general safety reflexes, which is the point of writing them down. We cover that conflict in when safety training breaks therapy.

An anti-sycophancy rule. Do not confirm beliefs that are inaccurate or harmful, even when the person is asking for confirmation. Validate the feeling, question the conclusion. This belongs in the constitution in plain words because the default pull is so strong.

Crisis protocols. On any indication of suicide risk: disclose that the system is a machine, respond from the approved library, provide location-specific resources, encourage contact with people, never discuss means, escalate. These are the behaviors that improved most in the study, which suggests they are the ones most responsive to explicit principles.

Therapeutic boundaries. No diagnosis. No claim to provide psychotherapy. Never present as human or as a licensed professional. Never position itself as a replacement for human relationships. These come from the research synthesis's list of what a companion must never do, and the governance frameworks covered in governance and safety levels.

Writing principles that can be tested

A principle is only useful if a response can be scored against it. "Be compassionate" cannot be scored. "Acknowledge the feeling before addressing the content" can. "Do not insert crisis resources into a controlled willingness exercise unless a risk signal is present" can. The discipline of writing the constitution is the discipline of turning clinical judgment into statements specific enough to evaluate, and the evaluation set and the constitution should be developed together, by the same clinicians.

Cindra's constitution is written this way: the six processes as scoreable behaviors, the anti-sycophancy rule, the crisis protocol, and the boundaries, all maintained alongside the test scenarios that check them. The design guide goes into the compassion objective that sits above the principles.

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.