Knowing what you don't know

The fastest way to lose a user's trust is to answer confidently and be wrong. In a general chat, a wrong answer is annoying. In a physical place — a clinic, a rental, a venue — it can send someone to the wrong door at the wrong time.

Confidence is a design decision

Language models don't come with a built-in sense of "I don't actually know this." Left alone, they'll produce a fluent answer for almost anything. So the willingness to not answer has to be designed in.

We'd rather the system say one of these than invent a detail:

• "That isn't set for this space yet — here's who can help."
• "I can tell you the general policy, but not the specifics for today."
• "I'm not sure. Let me point you to the front desk."

Grounded refusal beats confident guessing

A calibrated "I don't know" is more useful than a confident wrong answer — and far more trustworthy over time.

This is uncomfortable because it can feel like the product is doing less. But a place-based assistant earns trust by being reliable about the small set of things people actually rely on, and honest about the rest.

How we approach it

• Scope every claim. If a fact isn't grounded in the space, it isn't stated as fact.
• Prefer a handoff. When the system can't answer, route to a human path instead of improvising.
• Make uncertainty legible. Users should be able to tell the difference between "known" and "guessed."

Knowing what you don't know isn't a limitation we're working around. It's part of what makes an assistant safe to put in a real place.