AI Explainability

Existing in today’s world increasingly requires AI literacy, which should be afforded by better designed AI systems.

This project focuses on how we can better design AI systems, so their system decisions are understandable to the people they affect. This is an essential piece of responsible AI development and benefits society by reinforcing public understanding of AI’s impacts.

The keystone of this project is: explainability.withgoogle.com, a site which includes a rubric of 20+ items that AI builders, product managers, and regulators should seek to understand about AI systems. These are not pieces of the technical underpinnings of machine learning fundamentals, but most are choices made by deployers of AI, that have implications for the people and communities that use them.

By designing systems that meet as many as possible of these rubric items, AI systems become better explained, more transparent, and more accountable to people. To show this, we created a series of three interactive stories¹ built around a fictional sea navigation app, where people can follow along with our cast of animated characters as they see the explainability principles at work. Additionally, we have a set of activities teams can do to help understand how to implement the principles. We’ve run these worksheets at dozens of workshops inside and outside Google.

These workshops and our explainability consulting work with Google product teams have led to better product design: more transparency, clearer explanations and improved accountability to our users.

We’ve also developed a series of AI explainability lessons for middle schoolers called Discover AI in Daily Life and a series of Explainability Case Studies that were an early prototype of our workshops built more for policy makers around self-driving cars.

And all of this work builds on our understanding of how people think about AI systems, from public opinion surveys, qualitative interviews, and AI trust workshops we ran in coordination with Google’s People + AI Research (PAIR), beginning in 2019.

#Footnotes

1. Dolphin Sunrise, Whale Trail No. 5, and The Daysailer. ↩