Looking back, looking ahead: Humans, ethics, and AI

Abstract

Concerns about ethics of AI are older than AI itself. The phrase “artificial intelligence” was first used by McCarthy and colleagues in 1955 (McCarthy et al. 1955). However, in 1920, Capek already had published his science fiction play in which robots suffering abuse rebel against human tyranny (Capek 2004), and by 1942, Asimov had proposed his famous three “laws of robotics” about robots not harming humans, not harming other robots, and not harming themselves (Asimov 1950). During much of the last century, when AI was mostly confined to research laboratories, concerns about ethics of AI were mostly limited to futurist writers of fiction and fantasy. In this century, as AI has begun to penetrate almost all aspects of life, worries about AI ethics have started permeating mainstream media. In this column, I briefly examine three broad classes of ethical concerns about AI, and then highlight another concern that has not yet received as much attention. The first category of concerns about the ethics of AI—let us call this the superintelligence category—pertains to the fear that machines may one day become more intelligent than humans and harm human interests. In an extreme case of this type of concern, the fear is that AI agents may take over the world and then enslave or eliminate humans. As just one example, Bostrom (2014) imagines a futuristic world in which a superintelligent robot is asked to make paperclips and the robot pursues this goal until it consumes all of earth's resources, thereby endangering human existence. Some fears of superintelligent machines seem to derive from a mechanical “algorithmic view” of intelligence in which intelligence resides in an agent's brain and making a machine superintelligent awaits the invention of a master algorithm. However, intelligence in general is evolutionary and developmental, and human intelligence is also social and cultural. In particular, human intelligence is the result of numerous social interactions in which we learn from our parents, siblings and families, our teachers, peers and schools, our neighbors, friends, communities, and so forth. Human-level general intelligence in machines too will build on numerous social interactions with humans and other machines. Further, human intelligence is cultural: we learn about human goals, interests, values, norms, and meanings through our social interactions; in fact, these shared goals, interests, and values are a fundamental basis of our behaviors. Human-level general intelligence in machines tool will be based on similarly shared goals, norms, and meanings that derive from the machines’ interactions with humans, and they will be as fundamental a part of an intelligent machine's behaviors as its body and brain. From this social and cultural perspective on intelligence, the notion of a superintelligent machine that will produce paperclips until eternity and put human existence in danger seems a little odd. In contrast to the first category, the second set of concerns—the bias category—is not only more valid but also more urgent: data security and privacy as well as data and algorithmic bias and fairness. Concerns about data security and privacy are not specific to AI; they pertain to all of information technology. The field of cybersecurity and privacy seeks to address these concerns and therefore I will not explore them further. However, some of the worries about data and algorithmic bias and fairness directly pertain to AI and thus merit attention here. Dieterle, Holland, and Dede (in press) have developed a framework for understanding how biases from various sources feed on one another: (i) due to various factors such as age, health, skills, class, geography, and demographics, there is a citizenship divide in the society; (ii) the citizenship divide leads to different levels of access to hardware, software, and connectivity resulting in an access divide; (iii) the access divide leads to collection of different kinds and amounts of