When the AMA Describes the Problem, It Also Describes the Agent

5 minute read

The American Medical Association (AMA) published a short guide this year on using AI chatbots for health. It is one page. It puts five prompts and a list of warnings onto a single sheet. If you build health technology, it is worth reading twice, because it does two things at once. It tells patients how to use a general chatbot safely. And without meaning to, it describes the product that should exist instead.

The guidance is right

The structure the AMA chose is clean. Use AI to learn, to simplify, and to prepare for care. Five prompts cover the ground. Ask what could explain a set of symptoms and what would help tell those causes apart. Add your own context, your age, your situation, your goals. Ask for a plain explanation of a term or a diagnosis. Ask about treatment options and the benefits, risks, and side effects of each. Ask what questions to bring to your next visit.

Then come the limits, and they matter more than the prompts. AI cannot diagnose. It does not know your full history. It can sound confident even when it is wrong. The quality of what you get back depends on how you ask. Do not use it in an emergency. Do not let it stand in for your doctor. And be careful with your data, because a chatbot is not a clinic and does not carry a clinic's protections. Every line of that is correct. I would not change a word.

The tool was never built for the job

Here is the quiet problem inside the page. The AMA is writing safety instructions for a tool that was not designed for health at all. A general chatbot has no idea who you are when you open it. It forgets you between sessions. It has no obligation to keep you inside the safe lane of learning, and nothing stops it from drifting toward diagnosis if you push. It will happily absorb the personal details the AMA tells you to withhold. The burden of using it well falls entirely on the patient, who is often frightened, tired, and reading about their own illness for the first time.

That is a lot to ask of someone the week after a diagnosis. The five prompts assume a calm, informed user who remembers to add context, ask for trade-offs, and stop short of treating the answer as medical advice. Real people do not behave that way under stress. They ask the next obvious question, and the next, and a generic tool follows them wherever they go.

What a Personal Health Agent does differently

A Personal Health Agent is the same idea, made safe by design rather than by a warning label.

It starts from the same three jobs the AMA names. Learn. Simplify. Prepare for care. The difference is that the boundaries are built into the system, not printed on a sheet that the user has to remember. The agent is told, in its own instructions, that it explains but does not diagnose. When a question crosses that line, it does not improvise. It points back to a professional and, where the setup allows, helps the person get there.

It also holds context correctly. A general chatbot either forgets everything or quietly keeps too much. An agent built for this work can carry the parts of your situation that make its answers useful, your diagnosis, your medications, your questions for the next appointment, while treating that information as something to protect rather than something to harvest. The AMA tells patients to avoid sharing identifying details with a chatbot. The better answer is an agent where sharing those details is safe in the first place.

And it lives in the right moment. Most of the hard part of being a patient happens between consultations. The appointment is fifteen minutes. The weeks around it are when the confusion sits, when instructions are misread, when doses are missed, when the questions you meant to ask go unasked. That gap is where a Personal Health Agent earns its place. Not in the exam room, and not in an emergency, but in the ordinary days on either side of care.

Caregivers are half the story

The AMA page is written for the patient. In practice, a large share of health questions is typed by someone else. A daughter managing her father's medications. A partner is trying to understand a scan result before the next call. A friend who became a caregiver overnight and is now the one keeping track.

These people carry a real load and get almost no support. A caregiver facing agent, what we call a Care Partner, works on the same principles. It learns and simplifies, it does not diagnose, and it builds toward care rather than around it. The difference is the point of view. It helps the person who is helping someone else. That role has been invisible to most health technology, and it should not be.

Saying the limits out loud

It would be easy to read all of this as a claim that an agent fixes what a chatbot cannot. It does not, and pretending otherwise would repeat the exact mistake the AMA is warning against.

An agent still cannot diagnose, and it should never try. It can still be wrong, so its answers need to be the kind a person can check and bring to a clinician, not the kind that ends the conversation. It is still not the place for an emergency. The honest version of the pitch is narrow. A Personal Health Agent does not replace a doctor. It makes the time between doctors less lonely and less confusing, and it does that inside guardrails that hold even when the user forgets they are there.

The escalation path is not a feature we mention at the end. It is the spine of the design. Every agent should know what is not allowed to do, and what to do when it reaches that edge.

The page and the product

The AMA wrote a careful set of instructions for using the wrong tool well. That is useful, and people should follow it. But the deeper reading is that the safe way to use a general chatbot for health is to make it behave like something it is not. It should know you, protect you, stay inside learning and preparation, and hand you off to a professional at the right moment.

That thing has a name. It is a Personal Health Agent. The AMA described the rules. The work now is to build the tool that keeps them without being asked.