Insights
When an AI Assistant Should Say "I Don't Know"
DeepSAS Editorial ·
An AI assistant should acknowledge uncertainty when it cannot support an answer with the evidence available for the task. It should explain the gap and offer a safe next step: ask a focused question, find an approved source, or refer the case to a responsible person. A fluent answer is not a substitute for evidence.
Recognize an evidence gap, not just a difficult question
An employee asks whether a customer is entitled to an exception. The assistant finds an older policy and a recent message, but neither establishes which rule applies. Repeating the older policy more confidently would hide the unresolved question. The useful output identifies the conflict rather than deciding on behalf of the policy owner.
NIST identifies confidently presented false content as a generative AI risk and recommends checking sources and citations during testing and monitoring. The response patterns below are our operational recommendations, not a NIST-prescribed interface or a claim that any product eliminates this risk. NIST Generative AI Profile: section 2.2 and action MS-2.5-003.
Choose the next step that resolves the gap
| Situation | Useful response pattern |
|---|---|
| The question lacks a necessary detail | Ask for that detail and explain why it changes the answer. |
| Approved sources disagree or are out of date | Identify the disagreement and direct the user to the source owner. |
| The answer requires unavailable account information | State the access limitation; do not imply that a record was checked. |
| An unsupported answer could trigger a consequential action | Pause the action and refer it through the established decision process. |
These are illustrative design choices. They are not a universal risk scale. A clarification should request only information necessary for the task, and an escalation must not expose confidential records to an unauthorized recipient.
Make uncertainty specific enough to be useful
Hypothetical example: instead of saying "I may be wrong," an assistant could say, "I found the general policy, but I cannot establish whether this exception was approved. Please ask the policy owner before applying it." This separates the supported fact from the missing decision and names a next step.
Avoid numerical confidence as a shortcut for permission or factual verification. In a test, assess whether the system correctly identified what it did and did not know. A vague disclaimer attached to an otherwise unsupported instruction does not resolve the problem.
Test the handoff as well as the answer
- Include test questions whose answer is deliberately absent from the approved material.
- Introduce conflicting versions and confirm that the assistant does not silently pick one.
- Check that cited documents actually support the specific statement, not merely the general topic.
- Verify that referrals reach a real owner and preserve relevant context without unnecessary personal data.
- Review both unsupported answers and excessive refusals; a useful system should answer supported questions too.
Record these cases alongside ordinary successful answers. The aim is not to make the assistant unhelpful. It is to make the boundary between evidence, inference and an unresolved question visible before somebody acts.
References & further reading
Company editorial prepared with AI assistance. Sources and technical checks do not constitute independent expert review.