Most AIs are sure of everything.
Let each one send what it says it's sure about. Same 100 tweets.
sent, rightsent, wrongleft for a person
One message. Eight answers. One pass.
It scores every possible answer at once. The top score is how sure it is.
The first short test tweet in each language from a random sample, not picked to look good. Recorded answers.
It didn't speak Haitian Creole. Then it did.
SMS to Haiti's 4636 help line, 2010. How well each version finds the messages asking for rescue, medicine, water or food, and shelter.
Mean AUC: 1.0 is perfect, 0.5 is a coin flip. The SMS text is not shown here; that dataset states no licence.
Your turn at the desk.
Pick an AI and a rule. The red line on each slip is the safety line.