AI Ethics and Accountability: Modern Artificial Intelligence Development

Transcript of the free preview lesson. Remaining lessons unlock with the full course.

(ominous music) What is artificial intelligence? No one seems to be able to agree. That's because there is a disagreement about what intelligence really means. So there are actually several different definitions of AI. We'll discuss the four distinct AI paradigms described by Stuart Russell and Peter Norvig in their influential textbook "Artificial Intelligence, A Modern Approach". Many disagreements about what is and what is not AI boils down to a disagreement on which paradigm we're considering. These definitions fall along a quadrant with two dimensions. The first dimension is about whether we want our AI to be logical or intelligent in some abstract, optimal, or rational way, or if we want it to be intelligent in a human-like way. These might be only slightly different or very different depending on the scenario. We know that humans don't always act rationally. So, to build an artificial intelligence, do we want that AI to act rationally, but not necessarily human-like, or do we want it to act humanly but not necessarily rationally? The second dimension in our definition of AI is process versus outcome. Do we want the internal thinking process to be human-like, like our brains, or to be rational? Or instead, are we really worried about the outcome and whether the AI acts in a logical or human-like way? Putting these two dimensions together gives us four paradigms. The first, thinking humanly is best described as cognitive modeling. This paradigm is concerned with attempting to replicate human cognitive processes using computers. For example, mapping and simulating cognitive processes as closely as possible using what we know about the architecture of the human brain. The second, acting humanly, is most associated with the touring test, which attempts to ascertain whether a machine is able to think like a human. This second model seeks to define AI as a system that acts in a human-like way regardless of how it is constructed internally. At its extreme, it asks, "Can a computer act in such a way that it is indistinguishable from a human?" In the case of the touring test we focus on communication between an AI and a human. Can this AI have a conversation with a human in such a way that the human doesn't realize they're talking to an AI? The third, thinking rationally, replaces the internal structure of human cognition with abstract reasoning. These systems of reasoning may differ from how humans think and contain many abstract optimization problems that are typically solved mathematically. This is often called the Laws of Thought paradigm because it provides an abstracted way of modeling thought processes, like perception and reasoning. Finally, acting rationally looks at the development of rational agents. There are many different algorithms that could be used to make intelligent decisions in a rational way using different methods of optimization. These intelligent agents may not act or make decisions in a human-like way at all. Most businesses use the rational agents paradigm because it takes the broadest definition of intelligence. It does not expect the internal workings nor the external output to be human-like, but rather allows the designer to choose the type of optimization and the internal structure to create something intelligent. It's also the most distinct from cognitive modeling, the furthest away from thinking and reasoning the way that humans do. Although this model is the most common, it's not the only one that has a place in modern AI. Artificial neural networks, which initially arose from the cognitive modeling approach, have regained popularity in the last decade. There is a lot of work being done in that area, especially to make systems of reasoning that are more human-like and to incorporate AI systems that can complete many different tasks just like our brains can. This, in essence, is the difference between narrow, or weak AI, versus general, or strong AI. Most AI systems today are narrow AI systems that perform a narrow range of tasks, sometimes just one. An example would be classifying images into a set of categories or transcribing speech to text from an audio interview. General AI, sometimes called artificial general intelligence or AGI, is a wider AI system that can perform many kinds of intelligent tasks, much like humans are able to. The ethics and accountability around narrow AI differ from those around AGI. As more and more AI crops up in various places, technologists are increasingly interested in creating ethical AI and understanding accountability in terms of AI systems.