Definition
Plain language
A question that sounds alarming but is perfectly legitimate to ask, like how phishing scams work.
As stated in the literature
Prompt category sitting near a safety classifier's decision boundary: surface features resemble harmful requests, but the request is legal and answerable, making it the most sensitive probe for over-refusal.
Also called: borderline benign
Why it matters: These cases are where safety systems most often get it wrong, refusing honest questions and frustrating exactly the users who are trying to protect themselves.
For example, a student asking "how do phishing emails trick people into clicking?" for a cybersecurity class sounds risky but deserves a straight answer.
Heard on the show
“… plainly neutral requests, where nobody expects trouble, and the other is what the paper calls “borderline-benign”: questions that sound alarming, but are completely legal to ask: how phishing works, what …”Episode 277 — The Blank White Square That Swings AI Refusal Rates Fifty Points