Chapter 11. Fair Product Design and Deployment

For the most part, this book has focused on the data analyst’s and machine learning engineer’s dilemma. What is fair data to use? What are fair questions to ask? How can models be trained more fairly?

But the reality is that such work will make its way downstream into a product. That product may be what the word most commonly denotes, an actual consumer good or interface of some kind. Maybe it’s a game on which people will spend their time. Maybe it’s a report that will be made available for management to make decisions about resource allocation. Maybe it’s a SaaS that salespeople will go out and start selling to small government offices.

It is imperative that those upstream, producing code and models, think about how their work affects the products that are possible downstream. We have seen a trend of employees at large tech companies feeling deep concern and responsibility for the downstream uses of their work, as evinced by protests at Google, which have been successful, and Amazon, which have not been so succcesful when their employers have serviced contracts for controversial government uses of technology, such as immigration control or warfare.

It’s important that the people directly responsible for “product,” in whatever form that takes, develop a sense of responsibility and actionable guidelines for thinking about downstream uses of their product once they release it into the stream of commerce and consumption. Usually product ...

Get Practical Fairness now with O’Reilly online learning.

O’Reilly members experience live online training, plus books, videos, and digital content from 200+ publishers.