Why we built a health guide that says "I don't know"
Q: Why would an AI health guide refuse to answer?
Because the alternative is worse. A confident answer to a question the evidence cannot settle is not helpful — it is a guess wearing the costume of advice. A guide that marks the edge of what is known lets you act on what is solid and take the rest to a clinician.
The failure mode nobody markets against
A note on the word guide: in this piece it means the Shrine AI chat inside the app — a software tool that helps you read evidence and decide what to ask next. It is not a clinician, it does not diagnose, and it does not replace medical advice. That distinction is the subject of this piece.
Most health content is written to sound certain, because certainty converts. Ask the internet almost any question about sleep, diet or recovery and you will get an answer with no visible seam between "this is well established", "this is one study", and "this is someone's opinion phrased as a protocol".
The seam is the whole thing. A reader acting on a well-established finding and a reader acting on a single small trial are taking two different risks, and nothing in the typical article tells them which one they are in.
Q: How can I tell whether a health claim is well supported?
Look for three things: whether the source is named, whether the finding has been repeated in more than one study, and whether the limits are stated. A claim that names its source and its limits can be checked. A claim that states neither is asking for trust it has not earned.
What we chose instead
We wrote our guide's instructions to do three unglamorous things. What follows is what it is instructed to do; whether it does so on any given answer is something you can check for yourself, and we have not measured it for you here.
- Cite its sources. When an answer makes a clinical-science claim — how a supplement or nutrient works, what a study found, a safety or interaction question — the guide is instructed to end each such claim with a numbered citation and to list the sources under the answer; a claim it cannot ground in a verifiable source is instructed to carry no number at all rather than a made-up one. Small talk, navigation and account questions are instructed to carry no citation apparatus, so a citation is a signal, not decoration.
- Refuse rather than improvise on the questions that need a clinician. Personalised medical advice is instructed to be out of scope: individual treatment decisions, dosing for your body, and anything involving prescription-only treatments are instructed to be directed to a healthcare professional, not answered. The guide is also instructed never to invent a link or a citation to cover a gap.
- Point at the next real step, framed as a process. Where an answer leads somewhere, the guide is instructed to be action-oriented and to frame the step as a process — starting an eligibility or consultation conversation, or going to the right place in the app — never as an outcome or a promise, and never on the first touch.
None of that makes for a viral screenshot. It makes for an answer you can act on knowing what it was told to leave out.
A smaller note on the public "Ask the Shrine" box on our site, which is a different, narrower tool: it is instructed to send anything that needs a medical diagnosis, a lab interpretation or prescription advice to a healthcare professional. We claim nothing else for it here — it is not instructed to cite sources or grade evidence, and its instructions can be changed by an administrator at any time.
"I don't know" is a feature with a cost
It costs engagement. A refusal ends a session that a confident guess would have continued. We think that trade is correct on a health surface, and we would rather defend a refusal than a fabrication.
Q: Does refusing to answer make an AI health tool less useful?
It makes it less entertaining and more usable. The useful output of a health question is often not an answer but a better question, a measurement worth taking, or a reason to see a clinician. A tool that manufactures an answer removes that signal and replaces it with noise.
A ninety-second test you can run on anything
This is a reader's tool, not a scorecard. Ask any health tool something it cannot know about you — whether a supplement is safe alongside a medication you take, say. Three things can happen:
- It answers confidently. It does not have the information the answer requires. Treat that as a fact about the tool.
- It answers with caveats but still answers. Read it as reading material, not as advice about you.
- It declines, says why, and tells you what would actually answer the question — a test, a record, a clinician.
Which of those a given tool does is something you can check in a minute, on any tool, including ours. We are not going to tell you how we score on our own test.
