97 Things About Ethics Everyone in Data Science Should Know

Chapter 68. “All Models Are Wrong.” What Do We Do About It?

Miroslava Walekova

Senior Manager, Transformation Platform Ltd.

Machine learning will continue to transform every aspect of our lives: the way we interact with each other, the way we learn and develop, and the way we interact with society. Yet these systems will inadvertently break down every so often.

All models are approximations. Essentially, all models are wrong, but some are useful.

—George E. P. Box

In other words: no model, machine learning, or artificial intelligence solution can be right all the time. If we agree that failures cannot be avoided, then our main concern is to focus on the processes and controls that can effectively and efficiently minimize any adverse impact on individuals.

A machine learning governance framework has to cover solutions from idea inception to solution decommissioning, and it needs to:

1. Prevent solution problems by design

2. Rectify any issues in an expedited, transparent, and responsible manner

3. Improve the governance framework continuously

Let’s walk through each of these requirements.

1. Prevent

The effort to minimize adverse impacts starts with an internal assurance that a solution will adhere to principles of fairness.

Defining fairness, however, poses a number of challenges. Not only do individuals have different perceptions of what is fair, but there is also a great variety ...