Can you trust an AI assistant with health questions?

You can trust an AI assistant with a health question when you can check its answer against something: the letter it summarized, the result it quoted, the pharmacist it sent you to. An assistant that works from your own records and shows you where each answer came from makes that check quick.

I run a company that builds one of these assistants, Alo. The steps below are the ones I would want you to run on ours.

What is an AI assistant good for with health questions?

An AI assistant is good for health jobs where you can hold its answer up against a page. It can summarize a long discharge letter, list the dates in a stack of results, or turn a week of scattered notes into three questions for Thursday's appointment.

These jobs have something in common: the assistant is working from a document, and the document is still there afterward. What a term in a letter means, or which of three results is the most recent, is a question you can settle by looking. The letter is in your hand and you can read the line it summarized.

How do you get an AI answer about health that you can rely on?

Treat an AI answer about health the way MedlinePlus says to treat any health information: ask questions before you trust it, and discuss it with your health care provider before you rely on it. MedlinePlus wrote its questions for websites. This is my version for an assistant, in the order I would run it.

1. Give it the whole picture. Say who the question is about, and name the document or the date you mean. "What did Dad's cardiologist change in March?" gets a more useful answer than "what changed?"
2. Ask where the answer came from. Type: "Which document is that from, and where in it?" An answer about your father's medicines should come from his medication list or his discharge letter. If the assistant cannot point to a document, or its source is the open web, treat the answer as a question to take to someone.
3. Open the source and read the line yourself. Check whose name is on the page, the date including the year, every number digit by digit, and the unit after it.
4. Find out where your words go. Before you paste in a lab report, read the privacy policy for whether your data is sold or used for advertising, whether it is used to train models, and which other companies receive it. HIPAA usually does not cover an app you download yourself, so the policy is where the answer is.
5. Take it to a person before you act on it. MedlinePlus: "You may have found good information, but your provider can tell you whether it's good for you." On the phone to the clinic or the pharmacy, you can say: "An AI summary of Dad's discharge letter says one of his medications changed. Can you check that against what you have on file?"

What should you take to a doctor or pharmacist?

Deciding what is wrong with someone, or how urgent it is, belongs with a doctor or a nurse. Any change to a medicine belongs with the prescriber or the pharmacist.

An assistant helps you get ready for that conversation: the medication list, the dates, the questions in order. In an emergency, call 911.

What does Alo do?

Alo is the AI assistant in Lifelong, the family health app for iPhone. You can use it in the app or from Messages. Alo knows who Dad is, which is why it can tell you what his cardiologist said in March. Here is how it does on the steps above.

• Where an answer comes from. Alo answers from the health information in Lifelong that you are permitted to see, and your family's sharing permissions govern Alo exactly as they govern the app. Alo can also look things up on the web, so ask it which of the two an answer came from.
• Going back to the source. Lifelong shows where information came from, so you can go back to the source document.
• Changing a record. Alo can draft an entry for a health record, and your confirmation is what saves it. Every message and every proposed change is recorded in an audit trail.
• Where your doctor comes in. Alo leaves diagnosis and decisions about medication to your doctor and pharmacist. The Medical Disclaimer says where the line sits.
• Feedback. You can report any answer from inside the app.

How does Lifelong protect what you tell Alo?

A conversation with Alo is encrypted in transit. Our contracts restrict each model provider to processing on our instructions and ban training on what we send, and our model vendors are contractually required to keep no copy. Alo sends what the task requires rather than your whole record.

Lifelong does not train its own models on identifiable health data, and does not sell personal information or share it for advertising. The AI and Alo Disclosure lists what is sent for each feature.