We usually ask a language model for an answer. But when a system is unsure, continuing can be riskier than pausing. The question is not only how many answers it gets right; we should also ask when it refrains from answering.
The research behind this note examines abstention under different conditions and shows that the behavior can be changed by intervention. That does not mean the model understands uncertainty as a person does, or that it is generally trustworthy. It means “I don’t know” is a behavior worth measuring and designing for.
An unexplained refusal is not enough in a product. People need to know what is missing and what to do next: check a source, ask a specialist, or hand the task to a person. A model’s silence alone does not create safety; a clear workflow and human review matter too.