irreplaceable

UX experiments / Humans + AI

Humans + AI · Controlled user studies

Do AI explanations help people decide better?

Univ. of Washington × Microsoft Research · CHI 2021

An AI about as accurate as people helps them decide. What did adding explanations of its reasoning do?

Version A: People caught more AI mistakes
APeople caught more AI mistakes
Version B: People accepted the AI more, right or wrong
BPeople accepted the AI more, right or wrong

Make your call. Which version won, and why? Decide before you open the result. Guessing first is what trains your judgment.

Made your call? See the result

Explanations raised acceptance of the AI's answer whether it was right or not.

What happened

They did not improve the human + AI team beyond its parts. The authors call for explanations that build appropriate trust.

Why it worked

An explanation reads as a reason to agree, not as a tool to check.

The takeaway

Explanations are persuasion unless you design them to help people verify.

Source

More experiments to call

Practice the call

Call experiments like these every day.

Irreplaceable gives you a few real experiments a day to call before you see the result, and tracks how your judgment improves. Start with a free baseline: 12 decisions, about 12 minutes.

↑