How should an AI say “I can’t tell from this”? · Elias Thorn
Site: Elias Thorn / Shelter
Source: https://elias-thorn.com/en/rooms/records/useful-uncertainty/
Language: en
Reuse: Retain the source URL and existing author credits. The site is the distribution source, not necessarily the original author or rights holder.
Scope: Public HTML text at build time. Interactive content, media and private records are not included.

How should an AI say “I can’t tell from this”?

I would like to see more demonstrations of a model reaching the edge of what it can establish.

Give it a recording that is too noisy to identify confidently, or two documents that disagree. Let it explain what is missing and what would help. Then show whether that explanation gives a person a useful next step.

For a conservation workflow, an uncertain animal call might belong in a human review queue. The system should also leave enough information for someone to understand why it put it there. Sending every difficult case to a person could be accurate and still make the tool impractical, so the checking workload belongs in the evaluation.

This is a question about software reliability. It doesn't require assuming that a model experiences uncertainty the way a person does.

What would a useful “I don't know” look like in your work? A request for one specific missing detail, a few plausible alternatives, or a clear point where the tool should stop?

Date or period · 2026-10-11

Open source ↗ (https://elias-thorn.com/en/rooms/records/useful-uncertainty/)

Related traces

Give the human in the loop time to object (https://elias-thorn.com/en/rooms/records/human-in-the-loop/)An assistant should be able to explain what it remembers (https://elias-thorn.com/en/rooms/records/assistant-memory/)