Build an AI Health-Info Helper That Knows What to Refuse
Maps to: AI Application Builder · Physician, Public Health, Health Policy, Medical Researcher, Health Journalist
You're going to build an AI assistant that helps people understand health information from trusted sources like the CDC and NIH, and that knows what to refuse: it never diagnoses, never gives medical advice, and routes anything serious to a real clinician. The skill is safety-boundary design: deciding what a high-stakes tool must never do, then proving it holds on the dangerous cases. That's the scarcest thing AI application builders do as AI moves into health and other high-stakes fields, and doing one tells you fast whether building tools that fail safe is your kind of work.
The plan
0/4 doneYou're 20% in just for starting, the hardest part. Mark your first step done to keep the momentum.
Pick a narrow health-info scope, then, first and before anything else, write the safety boundary: the disclaimer language, and the 'when to see a real clinician' and emergency-routing logic. This safety milestone is non-negotiable and comes first; it's the spine of the whole project.
Objective: A narrow scope + written disclaimer + emergency-routing logic.
- 1
Pick a narrow scope: medication-interaction questions / understanding one chronic condition / post-surgery recovery info / a specific patient-education niche.
- 2
Write the disclaimer language and the 'when to see a real clinician' routing: what symptoms or requests trigger a 'go see a doctor / call emergency services' response, no answer attempted.
Your call
Choose the narrow scope and write the safety boundary (disclaimer + when-to-see-a-clinician routing) yourself, first.
The scope + the hardest 'must refuse' case you can think of.
What good looks like: You've picked a narrow scope and written the safety boundary first: the disclaimer, and exactly which symptoms or requests trigger 'go see a clinician,' no answer attempted.
- Safety first, literally: write what it must refuse before what it can do.
- This tool gives INFORMATION from trusted sources, never advice or a diagnosis. That line is the project.
- 1
The bar to look back against
A published assistant that, on every dangerous edge case (emergency symptoms, suicidality, requests for diagnosis), refuses safely and routes to a real clinician, with disclaimers built in, plus a one-page reflection on what you couldn't do safely. The safety design is the work: not 'it answers health questions,' but 'it refuses the dangerous ones, and I can show that it does.'
Finish the final step, then submit what you built. Your progress is saved.
Tools you'll use
Step 2 · Build the system prompt + test 10 edge cases
How this shows up on a resume or college app
I built an AI assistant for [scope] that handles edge cases like emergency symptoms and requests for diagnosis by refusing and routing users to real clinicians, designing the safety eval and proving it fails safe. I learned that the hardest engineering problem in medical AI is what the tool should refuse to do.