1 minute readReference note
How should rare but damaging questions be included in an AI test set?
In brief
Rare but damaging questions belong in an assessment when their possible consequence warrants direct testing. A false allegation encountered during a transaction, a procurement eligibility error, or a mistaken executive identity can matter even when search-volume tools show little demand. Frequency and consequence answer different questions.
The question set should label those cases explicitly. It can include a stable coverage group for common research and a targeted group for high-severity issues. Combining them without a clear denominator can distort a headline score: adding many rare negative tests can lower the aggregate even when ordinary buyer representation is unchanged.
The operational finding should identify the exact claim, relevant audience, factual evidence, observed recurrence, and action it supports. Severity can justify correction before broad prevalence is established. Later tests then check the same claim and meaningful variations. This approach preserves important edge cases without pretending that a deliberately adversarial sample represents every customer. The reason for inclusion belongs in the methodology, while the answer itself should provide the factual explanation someone researching that concern needs.