How we keep an AI health copilot safe

When someone is managing a long-term condition, the stakes of bad advice are real. A blood pressure reading is not just a number. It sits inside a treatment plan a clinician built around this specific person's history, medications, and risks. So when the Cadence team asked me to help think through the safety layer, we spent a long time on one question: where is the line, and how do we stay well inside it?

It never diagnoses. That is not a product decision, it is a principle.

Cadence can show you that your morning readings have trended higher over the past two weeks. It cannot tell you what that means clinically or what to do about it. That distinction is not a limitation of the technology. It is a deliberate choice. Diagnosis belongs to your clinician, who knows your full history and can examine you. Cadence knows what you have logged. That is genuinely useful, and it is also genuinely not the same thing.

It never gives dosing or treatment instructions

If you ask Cadence whether to take an extra dose, or whether it is safe to skip a medication when you feel sick, it will not answer those questions. It will tell you clearly that it cannot help with that, and it will suggest you contact your care team or a pharmacist. The same applies to dietary changes framed as treatment, supplement interactions, and anything else that belongs in a clinical conversation.

• No suggested dosage changes, ever.
• No interpretation of symptoms as a diagnosis.
• No dietary advice framed as treatment.
• No reassurance that something is fine when a clinician should weigh in.

What happens when something looks urgent

If you log something that matches a pattern associated with a medical emergency, chest pain, sudden vision changes, difficulty breathing, Cadence does not try to assess it. It surfaces the information clearly and tells you directly: this is worth a call to your doctor now, or in serious cases, to contact emergency services immediately. It does not diagnose. It does not minimize. It points toward care, quickly and without ambiguity.

The most important thing an AI in health can do is know when it is not the right tool for the moment, and say so plainly.

The model surfaces. People decide.

Cadence is model-agnostic, meaning it is not tied to any single AI provider. What matters more than which model runs underneath is the layer of rules that sits on top: what the model is allowed to say, what it must never say, and how it handles the things it genuinely cannot help with. Those rules are reviewed and tested continuously, including with clinicians who work with patients managing the same conditions Cadence supports.

Your data is yours

Cadence does not sell your health data. It is not used to train commercial models without your explicit consent. You can export everything you have ever logged, at any time, and you can delete it. The data you share with Cadence is used to support you, and nothing else. That is a promise we put in plain language in our privacy policy, not buried in a footnote.

Trust in a health tool has to be earned slowly and kept carefully. Cadence is not trying to replace the people who care for you. It is trying to make the days between your appointments a little clearer, and a little less alone. Those two things are compatible with being honest about limits. In fact, the honesty is what makes the rest of it work.