Design an AI Task With an Explicit Unknown Output

An extraction task becomes fragile when every field must contain a confident answer. Define an explicit Unknown state for information that the source does not establish. Keep Unknown separate from No and Not applicable.

Set up a practical example

For a fictional event note, “Parking is not mentioned” does not mean parking is unavailable. Ask the assistant to return the source wording, the extracted value and a status label. Specify how it should handle blank fields and conflicting statements.

Test the prompt with one clear fact, one omitted fact and one contradiction. Write the expected status for each before running the task. If the source says two different start times, the output should flag the conflict rather than silently choose the later sentence.

Verify before proceeding

Review whether the assistant preserves those distinctions in its summary. A table full of Unknown values may accurately reveal that the source is insufficient. The next action is obtaining information from the appropriate source, not making the table look complete.

When new evidence arrives, update the affected rows and record what resolved them. Keep previous uncertainty visible in the working notes when traceability matters. This method improves the representation of missing evidence; it does not guarantee the assistant will identify every unsupported claim.