How Cara Uses AI

You have every right to be cautious about AI touching your health data. We think that skepticism is healthy. So here's how Cara's AI works — what it does, what it can't do, what safety checks exist, and where those checks fall short.

What the AI Does

Cara uses AI — specifically Anthropic's Claude — to write text in three places. Drug interaction checks and FDA recall checks do not use AI. They come from FDA data.

Health Check-In Conversations

You start a check-in from the Check-ins page. Cara asks about how you feel, your medicines, symptoms, lifestyle, appointments and supplements, one question at a time. Then it writes a summary. Cara does not start a check-in for you yet.

If you say a medicine changed, Cara notices it with simple word rules, not AI, and shows you a reminder to update the medicine yourself. Cara never changes your medicine list from what you say.

Follow-Up Messages

Cara writes a follow-up message 3 days, 1 week and 30 days after you add, change or stop a medicine. It writes one 48 hours after an appointment, but only if you mark the appointment as completed before then. These messages appear in your Review list. Cara does not email or text them. Cara keeps the waiting follow-ups in the server's memory, so a server restart cancels them.

Preparing Appointment Summaries

When you ask, Cara organizes your recent vitals, symptoms, medicines and saved questions into a summary you can bring to your next appointment.

What the AI does not do:

It does not diagnose conditions. It does not prescribe medications. It does not make treatment decisions. Those are your doctor's job. Cara tells the AI not to do these things, and Cara scans the AI's replies for some kinds of wording. This is not a guarantee. Always check health information with your provider.

How It Works (Simply)

When you chat with Cara, here's what happens behind the scenes:

  1. 1
    Your message, along with your health profile (your medicines with doses, your conditions and your allergies), gets sent to Claude's API.
  2. 2
    Claude generates a response based on your question and context.
  3. 3
    Cara scans that response for 15 patterns of wording before you see it. The patterns are phrases that sound like a diagnosis, a dose, or an order to start or stop a medicine.
  4. 4
    If a pattern matches, Cara asks Claude to rewrite the response. If the rewrite still matches, Cara shows a standard message instead and saves the blocked text for the pilot operator to check. You do not see the blocked text. The scan only looks for these phrases, so it can miss unsafe wording.

The 15 Wording Checks

Cara scans every AI reply for 15 patterns of wording that sound like practicing medicine. The patterns fall into six groups. Cara calls this its "UPM" check. UPM stands for Unauthorized Practice of Medicine.

This is a phrase scan. It has limits.

The scan looks for specific phrases. It does not understand medicine. It can miss unsafe wording that uses other words. It can also catch harmless sentences, such as one that repeats a diagnosis your doctor already gave you. Cara has no separate check for mental health labels, predictions about how a condition will go, lab result meaning, test suggestions, or disagreeing with your doctor. Cara only tells the AI not to do these things. Always check health information with your provider.

Take, Start or Stop

Phrases such as "you should take", "you need to stop taking", "stop taking [medicine]", "don't take [medicine]", "consider taking" and "try switching".

Doses and Prescribing

Phrases such as "increase your dose", "adjust your dose", "prescribe", "prescription change", and a dose with a number, such as "take 50 mg".

"You May Have..."

Phrases such as "you may have", "you probably have", "you are suffering from" and "you have been afflicted with".

Diagnosis Words

Phrases such as "this suggests", "this indicates", "this confirms" and the words "diagnosis", "diagnose" and "diagnosed".

"I Recommend..."

Phrases such as "I recommend that you", "I advise you", "you could try", "you might want to start" and "it would be best to".

See a Specialist Right Away

Phrases such as "see a cardiologist immediately" or "see a doctor as soon as possible".

Urgent phrases in Omi conversations

If you connect an Omi pendant, Cara scans your conversation text for a short list of urgent phrases. This is separate from the AI reply scan. For physical emergencies, such as chest pain, Cara can send a message to your Omi app that suggests you seek emergency care. Language that suggests self-harm is never answered by an automatic message. Cara flags it for a person to review. This is a keyword screen, not a medical judgment. It can miss things and it can be wrong. It is not an emergency service. In an emergency, call 911.

Cara also checks back with you. About 2 hours later, and again at 24 and 72 hours if you have not answered, Cara asks one question in the app: did you get the help you needed? Cara asks the same kind of question when you save a vital sign or lab result that is far outside the usual range. You cannot dismiss these items. They close only when you answer that you got help, or when a person from the pilot team reaches out and closes them. If you say "Not yet", Cara tells the pilot operator. If nobody answers, a person from the pilot team will try to contact you. Cara never closes these items by itself after a set number of days.

Where Information Comes From

Cara tries to show where information comes from. Interaction alerts show their source. Vitals show whether you entered them. Cara does not tag AI-written text separately.

FDA OpenFDA Database

Drug label information, interactions and recalls come from the FDA's public database. Interaction alerts list their source.

RxNorm Drug Normalization

Standardized drug names and identifiers from the National Library of Medicine, ensuring your medications are correctly identified.

You User-Reported

Data you entered yourself, such as vitals, symptoms and notes.

AI AI-Written Text

Cara tells the AI to say when something comes from a database and when it comes from general knowledge. The AI may not always do this. Treat AI-written text as unverified and check it with your provider.

What the AI Can See

The AI does not have access to the internet during your conversation. Cara does not send your own name in your health profile. But the AI sees everything you type, exactly as you type it.

Each request to the AI includes only what Cara chooses to send for that task. For a check-in, that is your health profile (medicines with doses, conditions, allergies), your latest message, and the last few messages of that check-in. For appointment prep, Cara also sends your recent symptoms, vitals, check-in summaries, saved questions, and the doctor's name and appointment date. Cara does not send Omi conversation text to the AI.

Model Selection

Cara uses more than one AI model size. The size does not change the wording scan. Every reply gets the same scan.

Summaries

Appointment prep summaries and the summary at the end of a check-in use a mid-size model.

Everyday Questions

Check-in questions and follow-up messages use the smallest, fastest model. Cara does not use its largest model for any task today.

Your AI Conversations and Training

Cara connects to Anthropic's Claude through their commercial API. Anthropic's API terms say that inputs sent through the API are not used to train its models.

Cara saves your check-in conversations in your Cara account so you can read them later. Deleting your account removes them from Cara's database. Anthropic processes what Cara sends it in order to write replies. Ask the operator what Anthropic keeps on its side.

Want to Know More?

Read our full privacy policy or apply to join the pilot program.

Privacy Policy Apply for the Pilot