Choose which mistake your AI makes
Choose which mistake your AI makes
Choose which mistake your AI makes
The short answer
Every AI classifier will be wrong sometimes. Decide whether a false alarm or a miss costs people more, then make that mistake cheap to recover from.
Every AI classifier will be wrong sometimes. Decide whether a false alarm or a miss costs people more, then make that mistake cheap to recover from.
Illustrative concept, not client work or measured results.
Price both errors before you tune
Picture an assistant that flags meeting requests in email. It can fail two ways: it flags a newsletter, or it misses a real request. One accuracy number hides that choice.
Kocielnik, Amershi and Bennett tested a scheduling assistant like this. Two versions had the same 50% accuracy. The one tuned to avoid misses was rated more accurate and more acceptable than the one tuned to avoid false alarms. The authors suggest recovery cost explains part of it: an extra highlight can be ignored, while a miss means rereading and scheduling by hand.
It was one email task with crowdsourced participants. Treat it as a question to test, not a rule. Still, it shows that recovery cost is a design decision, not just a model setting.
Price both errors before you tune
Picture an assistant that flags meeting requests in email. It can fail two ways: it flags a newsletter, or it misses a real request. One accuracy number hides that choice.
Kocielnik, Amershi and Bennett tested a scheduling assistant like this. Two versions had the same 50% accuracy. The one tuned to avoid misses was rated more accurate and more acceptable than the one tuned to avoid false alarms. The authors suggest recovery cost explains part of it: an extra highlight can be ignored, while a miss means rereading and scheduling by hand.
It was one email task with crowdsourced participants. Treat it as a question to test, not a rule. Still, it shows that recovery cost is a design decision, not just a model setting.
Make the chosen mistake cheap
Google PAIR warns that weighing false positives and false negatives equally by default rarely matches real consequences. So choose deliberately, then design for the error you chose.
In this illustrative concept, the assistant leans toward catching more. A wrong flag clears in one tap, the email stays untouched, and Undo stays close. The setting says the trade-off plainly: some flags will be wrong.
Make the chosen mistake cheap
Google PAIR warns that weighing false positives and false negatives equally by default rarely matches real consequences. So choose deliberately, then design for the error you chose.
In this illustrative concept, the assistant leans toward catching more. A wrong flag clears in one tap, the email stays untouched, and Undo stays close. The setting says the trade-off plainly: some flags will be wrong.
Bring three questions to your next design review. Which error did we choose? Who pays for it? How many taps does it take to recover?
Bring three questions to your next design review. Which error did we choose? Who pays for it? How many taps does it take to recover?