Why the assistant only cites sources you could look up yourself

A made-up phone number is a small error for a chatbot. It is a terrible night for a caregiver who dials it at 9 p.m. with a real problem. This post explains the single design decision that does more for trustworthy answers than anything else we built: our assistant cannot cite the open web.

First, the problem in plain terms. A general AI chatbot writes by predicting what text should come next, and "plausible" is not the same as "true." Ask about Medicare coverage and it can produce a rule that sounds exactly like a Medicare rule but was never one — or name an agency that does not exist. The industry calls this hallucination. In caregiving, where answers turn into phone calls and decisions, it is the failure mode that matters most.

A closed list of 27 sources

When the assistant needs information beyond your profile — what a benefit covers, how to find local services, what a diagnosis typically means for daily life — it runs a live search that is restricted to a curated list of 27 sources. Not "prefers." Restricted. A forum post or a content-farm article cannot enter the answer, because the search never sees them.

How the list was chosen

The criteria are boring on purpose. Government programs come first, because for benefits and eligibility they are the authority — Medicare rules from Medicare.gov, not from a site summarizing it. Medical information comes from major research institutions and established disease organizations. Practical caregiving guidance comes from organizations that have served family caregivers for decades. What never makes the list: personal blogs, discussion forums, content farms built for ad clicks, and any site whose business is selling the thing it recommends. The list is deliberately short — every addition has to make answers more checkable, not just more numerous.

Every recommendation carries a link

Answers that draw on research include citations you can click — a real page on Medicare.gov, a National Institute on Aging guide, an Eldercare Locator lookup. This is not decoration. It changes your relationship with the answer: you are not asked to trust the AI, you are handed the source. Bring the link to the discharge planner or the family meeting instead of "the AI said so."

One honest caveat: a citation makes an answer checkable, not medical. Even a perfectly sourced answer about symptoms is still research to bring to a clinician, and the assistant is built to say so — see the guardrails that keep it in its lane.

Standing rules against making things up

The assistant's instructions include a hard rule: never fabricate agency names, phone numbers, addresses, or certifications. When it cannot verify something, it is built to say so and point you to a human who can — behavior we verify against a 118-scenario test suite, described in how we test the AI before it talks to your family. For local services, it routes you to your Area Agency on Aging or the Eldercare Locator — the front door for aging services in every U.S. county — rather than guessing at a listing.

Why not use the whole internet?

It is a real trade-off, and we made it deliberately. Somewhere on the open web is a niche blog post with a genuinely clever tip our closed list will miss. But you can find that with a regular search anytime, and you can judge it with your own eyes when you do. What you cannot do at 9 p.m. on a hard day is fact-check a confident paragraph of AI text line by line. So we chose verifiability over breadth: less internet, more right.

This is one of the four big differences from a general chatbot — the others are covered in why not just use ChatGPT.