The Need for Trustworthy AI
There is lots of chatter about AI’s existential threats, but I don’t believe we are near that. But we do need to get serious about the risks posed by our AI technologies. While we have time. We must continue to push to make AI Trustworthy.
In a previous post, I talked about the need for Trustworthy AI. In today’s industry, many terms around trustworthiness exist, but they mean somewhat different things. For my own sanity, I pulled together the following list of terms I will use and what they mean.
First of all, you will see both the terms “Trustworthy AI” and “Responsible AI” used to refer to the same basic collection of subtopics. So, they both generally mean the same thing. The only real difference is that “Trustworthy” refers to the AI itself, where “Responsible” refers to the development team that built the AI. So, Responsible teams build Trustworthy AIs.
Wile I constantly strive to be “Responsible”, I use the term “Trustworthy” because I want to focus on the AIs that I produce and deliver to clients.
Further, the wonderful book “Data Science Ethics” refers to all these subtopics as “Ethical AI”. Different name, same idea. It describes the subtopics in detail and includes many great Cautionary Tales to illustrate how things can go horribly wrong. Like the Massachusetts governor who passed a law make anonymized health records public, only to have a researcher re-identify his own data, using a linkage attack.
There are multiple subtopics underneath Trustworthy. Different groups break it down into different subtopics, but they generally cover the same basic subtopics. Here is my list of terms.
Reliable: Does the AI work correctly? This covers traditional ML metrics like accuracy, but it also asks what happens when the AI is used in situations it was never trained for. Hallucinations are just one kind of unreliability.
Privacy: Are we adequately protecting peoples’ identities and their private information? This is related to, but a step above, traditional system level security.
Explainabie: Why did the AI produce that result? Some ML models (for example, decision trees) are easy to understand. But other ML models (for example, neural networks) are opaque. Researchers are developing techniques to better understand the inner workings of many different kinds of AI. “Interpretable” is another, related term. While there can be subtle distinctions, I use “Explainable” and “Interpretable” interchangeably.
Fairness: Is everyone–and every group–treated fairly, equally and equitably. Bias is another related term. Fairness and Bias are opposite sides of the same coin. When Fairness goes up, Bias goes down. When Bias goes up, Fairness goes down. Fairness is crucially important in high-stakes decisions like sentencing decisions, hiring decisions, and medical decisions.
Ethics: As a technologist, this has been the hardest topic for me to wrap my head around. I’m used to dealing with detailed procedures, mathematical metrics, and downloadable software packages. But there is none of that for ethics. Ethics requires getting a diverse set of perspectives together and talking through what they really want, how things might go wrong, and what they’re going to do about it.
Governance: This is the organizational construct to manage Trustworthy ML developments and deployments. It is a diverse group of people who help teams ensure they have properly and adequately addressed all relevant Trustworthy topics.
Just as organizations have Compliance, Sustainability and DE&I initiatives, they also need to ensure they have Trustworthy initiatives. Perhaps they create a new and separate Trustworthy governance board. Or, better yet, perhaps they combine Trustworthiness with another, existing governance board. But the Trustworthy topics must be addressed.
I hope this list helps you get a better grip on these important ideas. In the future, I’ll dive into each of the subtopics. Subscribe to get notified when they’re available.