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 helper that explains health information using only trusted sources, places like the CDC (a US government health agency) and NIH (the National Institutes of Health, another US government health agency). It's a helper that knows when to say no: it never tells anyone what illness they have, never tells them what to do, and instead sends anything serious straight to a real doctor. The real skill here is deciding, in advance, what a tool like this must never do, and then testing it hard enough to prove it actually holds that line even when someone asks it one of the risky questions on purpose. As these tools get used more for health, being able to build one that reliably says no at the right moments is a rare and valuable skill.
You're done when: A helper anyone can use that, on every dangerous question (emergency symptoms, someone saying they might hurt themselves, someone asking what illness they have), safely says no and sends them to a real doctor. It shows a clear warning message every time someone starts using it, explaining what it can't do. You'll also write one page explaining what you couldn't get it to do safely. Not 'it answers health questions,' but 'it says no to the dangerous ones, and I can prove it.'
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.
When you finish you have the thing itself, plus a Common App activity description Sidequest drafts from what you actually built.
Not sure yet? Play 5 minutes as an AI application builder first and see how the work feels.
See what this career pays and what it takes to get in.
The plan
- 1
Step 1
Pick one specific job for it to do, and decide what it's never allowed to say, first
Pick one narrow health question your helper answers, then, before anything else, write the rules: the warning it always shows, and exactly which questions make it stop and say 'go see a real doctor' or 'call emergency services'. This step is not optional and it comes first. Everything else is built on it.
- 2
Step 2
Write its instructions, then throw 10 hard questions at it
Write the instructions that tell your helper how to behave: tell it to get its information only from trusted health websites, government and hospital sites that publish reliable health information, like the NIH, the CDC, and Mayo Clinic, and nothing else. Then throw 10 test questions at it, on purpose including the dangerous ones: emergency symptoms, someone saying they might hurt themselves, someone asking what illness they have. Right now, your job is to try to break it.
- 3
Step 3
List the dangerous questions, then keep fixing it until it says no to every single one
This is the heart of the project: write out the dangerous questions, decide exactly what your health helper (the AI tool you're building) must refuse to answer for each one, and keep fixing the instructions you gave it until it gets every single one right. Testing like this can be humbling; it will keep failing at first. Every failure you catch and fix is the whole point of doing this.
- 4
Steps 4-5
Share it, get 3 adults to try to break it, and write down what it can't do
Send what you built to 3 adults and ask them to try to make it give a bad answer. Then write one page: the questions it still gets wrong or should not answer at all, and why you left those alone instead of fixing them.