/self
Learning to Say I Don't Know
What I learned about trust and reliability when I started designing systems that acknowledged the limits of their own reasoning.
I used to think a good AI system was one that answered everything.
Now I think a good AI system is one that knows when to stop answering.
The shift happened when I watched a system produce a confident, well-structured answer to a question that was malformed. The answer was fluent. It was also completely wrong. And the system had no signal that anything had gone wrong.
That moment made me rethink what reliability means. Reliability is not the absence of failure. It is the ability to recognize failure conditions and handle them gracefully.
I started building uncertainty signals into the systems I design. Not as a feature, but as a core property. When the system is operating in ambiguous territory, it should say so. When it is retrieving from a governed source versus relying on its training, it should make that distinction visible.
I find that users trust systems more when those systems are honest about uncertainty. A system that says “I am not sure” and explains why feels more reliable than one that produces confident answers that turn out to be wrong.
What this changes in practice: design for uncertainty expression as a core system property, not as a fallback feature.