The White House recently unveiled the AI Bill of Rights. What is this, why is it needed, and what does it do? In this post, I provide some perspective — from a technology and business standpoint, about this document and what it can or should mean for businesses.
Context — why is this needed
First, some context. As many people know, AI is now deployed or being deployed in virtually all business contexts, from finance (think credit card approvals) to healthcare (think disease diagnosis/risk assessments) and beyond. While many advanced technologies stay far away from consumers — when was the last time you thought about how the latest database research was used by your hospital to treat you? — AI is not like that. AI advances touch humans directly — whether it is in using human information, making decisions that affect humans, or both. Plus, AI is pervasive — now anyone from a large bank to a high school student can leverage the most advanced AIs known to humanity, and then put the resulting application in front of anyone. How do we make sure this ferocious pace of technological advancement is safe?
Enter AI Ethics
AI Ethics is the field of AI that focuses on the ethical application of AI — particularly as it relates to humans and society. AI Ethics covers areas such as AI bias — ensuring that AIs treat all humans fairly, AI and Privacy — ensuring that people can understand and control how their information is used etc. AI Ethics is a critical field as explained here. Now we can get to the AI Bill of Rights — which is a prototype design for AI Ethics in practice. The AI Bill of Rights outlines the government’s view on what human rights should be protected by organizations building and deploying AI.
What is in the AI Bill of Rights?
A detailed document on the AI Bill of Rights can be found here. The document outlines five basic rights. I have listed them below:
- Safety. The key element here is that automated systems can (and do!) make mistakes. In AI — these mistakes can occur in many ways — see the article here about how COVID-19 broke many AIs around the globe. While not all mistakes can be foreseen, operational ML techniques (MLOps) can be used to detect and mitigate AI mistakes before they cause further damage.
- Privacy. AI thrives on information. Combined with the pervasiveness of sensors, video cameras, and records of online activity, it is now possible for vast amounts of personal information to be used by organizations without the individual’s awareness. This element focuses on the need for individuals to have methods to access, understand, and control the way their personal information is used.
- Fairness. AIs learn patterns from data. Without suitable data scrutiny, AIs can (and do!) learn biases and can become unequal in their treatment. This element focuses on the need to design and test AIs for fairness.
- Explanation. The Privacy right focuses on the need for individuals to be able to understand what information about them is used by an algorithm. The Explanation right is complementary — it states that the individual also has the right to understand how the algorithms use the data they are allowed to use. For example — if an individual has agreed to let a bank use their personal data (per the privacy requirement), the explanation requirement will show them if their age, gender, or any other information was used to determine a loan interest rate for them.
- Alternatives. This element focuses on the need to give individuals choices. The choice can be to opt-out of a system that is making automated decisions or have access to solutions or people to remedy problems caused by an automated system.
The AI Bill of Rights, in my opinion, outlines an interconnected set of principles that can be applied at all stages of the AI Lifecycle via a combination of AI Ethics and MLOps techniques. How it can be applied is very domain specific — for example, applications in healthcare will impose different constraints on human privacy than applications in web-based retail. However, the principles hold across all domains. It is worthwhile to examine each of these pillars and understand how it should fit into the operational AI practices in your organization.