Turn AI confidence into a next step
Turn AI confidence into a next step
Turn AI confidence into a next step
The short answer
Show confidence only where it changes what someone should do. Use a few plain levels, give each one a clear action, and offer a short list of options when the AI is unsure.
Show confidence only where it changes what someone should do. Use a few plain levels, give each one a clear action, and offer a short list of options when the AI is unsure.
Illustrative concept, not client work or measured results.
A score can calibrate trust, not fix decisions
Zhang, Liao and Bellamy had crowd workers predict incomes with help from an AI. When the AI's confidence was shown, phrased as how often it was right on similar cases, people leaned on it more when it was confident. A per-prediction explanation chart did not have that effect.
But calibrated trust did not make the joint decisions more accurate. The authors point to one reason: cases the AI found hard were hard for people too. And a score is only honest if the model's probabilities are well calibrated.
So confidence is not a trust badge. It is a cue for when to check.
A score can calibrate trust, not fix decisions
Zhang, Liao and Bellamy had crowd workers predict incomes with help from an AI. When the AI's confidence was shown, phrased as how often it was right on similar cases, people leaned on it more when it was confident. A per-prediction explanation chart did not have that effect.
But calibrated trust did not make the joint decisions more accurate. The authors point to one reason: cases the AI found hard were hard for people too. And a score is only honest if the model's probabilities are well calibrated.
So confidence is not a trust badge. It is a cue for when to check.
Give every level an instruction
Google PAIR warns that numeric confidence assumes people understand probability, and asks what anyone should do at 85.8% versus 87%. One alternative it describes: a few categories, each with a clear action. When confidence is low, showing several likely options prompts people to use their own judgment.
In this illustrative concept, a receipt assistant reads a lunch receipt. Fields it is sure about are filled in and stay editable. The category it is unsure about shows three guesses and waits for a choice before saving.
Give every level an instruction
Google PAIR warns that numeric confidence assumes people understand probability, and asks what anyone should do at 85.8% versus 87%. One alternative it describes: a few categories, each with a clear action. When confidence is low, showing several likely options prompts people to use their own judgment.
In this illustrative concept, a receipt assistant reads a lunch receipt. Fields it is sure about are filled in and stay editable. The category it is unsure about shows three guesses and waits for a choice before saving.
Before you add a percentage
Ask your team three questions. What will someone do differently at each level? Is the score calibrated against real outcomes? Have we tested the display with the people who will rely on it?
Before you add a percentage
Ask your team three questions. What will someone do differently at each level? Is the score calibrated against real outcomes? Have we tested the display with the people who will rely on it?