Responsible AI, Quantified

Hi, I’m Geoff Schaefer. I lead our AI ethics and safety work at Booz Allen Hamilton. Today, I want to talk about the future of AI ethics. The field is still in its earliest, most formative days. This gives us a unique opportunity to shape its evolution—and I think there’s a compelling reason to rethink how we apply the tools, techniques, and concepts in AI ethics today. Before we dive in, let’s set the stage. The field of AI ethics is in a paradoxical state. On the one hand, it has never been healthier. It is one of the fastest-growing subfields of AI. There has been a profusion of ethical guidance in the form of frameworks, principles, and toolkits from organizations big and small, public and private. The issues of bias, fairness, transparency, and other focal points have been cemented as domains in their own right. And it is now considered impolitic to establish an AI practice without an associated body of work in responsible AI. As we think about the future of AI ethics, the field should emphasize three questions: First, what is the maximal good an AI system can do? Second, what are the potential risks in its design? And third, how can we mitigate those risks so that the maximal good can be achieved? The order of these questions is intentional, as they shift our focus from harms to happiness, and from failure to flourishing. This will help us open up new missions and needs for AI ethics to support. After all, ethics was never about compliance. Nor was it simply about the difference between right and wrong. Rather, it provided the overriding question of philosophy in ancient times: How can we be happy and flourish? 

We discussed before that the most prominent questions in AI ethics are focused on mitigating risk and avoiding harm. Examples include ensuring the privacy of users, reducing bias in training data, and maximizing the explainability of a system’s outputs. But if we shift our attention from focusing exclusively on questions of harm to the concern of human flourishing, we’re able to expand our ethical calculus in helpful ways. And this has important implications for thinking about the types of AI systems that we should build and how we should build them. Let’s demonstrate this by looking at a case study. Some of the most promising applications of AI can be found in healthcare, including precision medicine and early-stage diagnostics. Yet, two of the most common AI ethics questions in this space are: How can we protect patient privacy? And how can we ensure our training data isn’t biased? Well, both of these questions are clearly very important; neither one of them directly addresses the principal goal of healthcare: reducing human suffering. To better align with this goal, what if we asked questions such as: How can we better treat rare diseases? Or, how can we prevent a patient from getting sick in the first place? These questions drive much of the science behind our most promising medical advances, but they rarely factor into our ethical calculus of the role AI plays in these same advancements. The AlphaFold AI system from DeepMind is a perfect example. This algorithm has been widely touted as having cracked biology’s 50-year challenge of protein folding. Proteins are the building blocks of life, but deciphering their structure—a process called “folding”—is notoriously difficult, taking months or years to achieve, if at all. But understanding protein folding is essential for everything from developing medicine to understanding antibiotic resistance. And while there are hundreds of millions of proteins known to science, AlphaFold has already folded over 200 million of them. AlphaFold has rightly garnered the attention and admiration of the AI community and beyond. But we don’t celebrate it as an ethical algorithm. Why not? Given the criticality of protein folding to modern medicine, it stands to become one of the biggest drivers of human flourishing in history. And it will do so in many different ways, from the elimination of debilitating disease to the strengthening of our immune systems. A more holistic and future-focused approach to AI ethics, then, would encourage the development of similar types of AI systems across a range of critical sectors and applications that impact human lives. In the case of AlphaFold and similar types of AI systems, this might include increasing access to their powerful technology to ensure they benefit all of society. Coming full circle here, what does it mean to live a good life? How can AI help us flourish? These are questions that AI ethicists should make central to their work. We should consider an AI system’s potential benefits and its potential risks in concert with one another. In fact, a more robust—and historically accurate—ethical calculus will focus on the net good that an AI system will generate over its lifespan.