A checklist for using a health AI safely, and for building one that earns your trust

After weeks of reading about scribes, answer engines, chart copilots, patient assistants, regulators and data, the same seven questions keep coming back. They work two ways: as a checklist if you use a health AI, and as a spec if you build one.

1. Does it show where the answer came from?
A link is not proof, but a source you can open is the minimum. If nothing is shown, treat the answer as a guess.

2. Does it tell you when it does not know?
The honest answer to "I don't have a source for that" is a good sign. Tabibu labels general-knowledge answers as such and says when it cannot find something.

3. Does it say what it will not do?
Tabibu does not diagnose, call a result normal, or change a dose. A tool with no stated limits has not thought about them.

4. Who pays for it?
Ads, subscriptions, grants and employers all shape a product. Look for a tool that says which one, and how the answer is kept separate from it.

5. What happens to your data?
Ask who sees it, why, and for how long. For my own documents feature the file is not stored, the answer is kept about 48 hours, and keeping it for 30 days is opt-in.

6. Is there a way to a human?
For anything serious, a good tool points you to a clinician or emergency numbers. Tabibu's emergency check is rules, not a model, and no clinician has reviewed those rules yet.

7. Are the limits published, including the bad numbers?
My own HealthBench re-run scored 0.426 against 0.522 for the bare model, on 150 conversations, not the official leaderboard. I would rather you read that than a chart with no method.

For the builders

If you build health AI, copy the list. Write each answer as a sentence on your site and as a test in your code. The ones you cannot answer in one sentence are your next tickets.

Limits and test results for my own project are at tabibu.health/trust.

I post short videos in this series on LinkedIn and TikTok; come say hi.

General information, not medical advice.