When Do We Trust Machines?

We propose a framework of "trust heatmap", show how the trust in machines depends on two key elements: their error rate and the costs of mistakes, and examine the automation frontier.

By Vasant Dhar, NYU.

In this presentation, I answer the question "when do we trust machines?" by walking through various situations in our everyday lives -- investing, playing sports, riding in driverless cars, and using social media platforms -- to encourage us to question the faith we put in technology.

In answer to this question, I use a "trust heatmap" in order to illustrate how the answer depends on two key elements:
  • how often machines make mistakes and
  • the costs or consequences of these mistakes.
Vdhar Ai Heatmap
Fig. 1: The AI Heat Map: Predictability vs Cost per Error

I show that automation occurs when problems cross an "automation frontier" when the risks are sufficiently reduced either through better data and algorithms, or because of regulation or an expression of our preferences. When used in this way, the heatmap can be used to predict what kinds of tasks are currently amenable to automation and those where humans should maintain control.

This ideas are presented in TEDx talk, entitled "When Do We Trust Machines?"

For more in-depth discussion, see HBR article

When to Trust Robots with Decisions, and When Not To (may require registration)

Bio: Vasant Dhar (@ProfDhar) is a Professor at the NYU Center for Data Science and the Stern School of Business and previous editor of Big Data Journal. His research addresses the question: when do computers make better decisions than humans?