Imagine two friends. Ask the first for a restaurant recommendation and she always has one—confidently, instantly, for every city on Earth. Ask the second and he occasionally says, “I haven't been there, but I can help you compare these three.”
The first friend is more impressive for about five minutes.
Many AI products are designed like that first friend. The interface contains an input and an answer. There is no natural place for “the documents do not contain this,” “the image is unreadable,” or “these sources disagree.” So the system converts missing evidence into a paragraph. The paragraph even has bullet points. Very reassuring.
The always-answer machine
Accuracy benchmarks usually ask whether the answer was right. Product interfaces quietly ask a second question: did the model answer at all? If an empty response feels broken, teams optimize it away. The model learns that guessing is better product behavior than stopping.
Research on uncertainty-based abstention points in the opposite direction: when uncertainty is measured appropriately, abstaining can improve reliability and reduce hallucinations. The difficult word there is “appropriately.” Asking a model “Are you sure?” is not a calibration system.
Abstention is a product state
“I don't know” is only annoying when it is a dead end. A useful abstention should name the missing piece and offer the smallest next move.
I found the cancellation policy, but not the date this plan started. If you share the first page of the agreement, I can calculate the window.
That response is not less capable than a guessed date. It is doing a different job: protecting the user from invisible uncertainty while keeping the task moving.
Make uncertainty useful
I like three kinds of exits:
- Missing evidence: say exactly what could not be found.
- Conflicting evidence: show the disagreement instead of averaging it into nonsense.
- Unsafe action: prepare a preview and hand control back.
None require showing a fake confidence percentage. “63% confident” looks scientific and gives most people no sensible action. Concrete uncertainty is easier to use: two invoices disagree; the source is older than the policy; the requested customer is ambiguous.
Score the silence
Evaluation should reward correct abstention and punish unnecessary abstention. Otherwise the safest model wins by refusing everything, which is just a very expensive error message.
Measure coverage alongside error: how many cases the system handles, and how reliable it is on those cases. Then inspect the handoff. Did the assistant preserve useful work, ask a good question, and explain why it stopped?
Intelligence is not answering every question. Sometimes it is recognizing the shape of the hole in the evidence—and resisting the urge to decorate it.