Thu, 02/10/2022
The Nebraska Governance and Technology Center had the pleasure of hosting Brian Christian, whose work on artificial intelligence has appeared in publications ranging from The New Yorker, The Atlantic, and the Guardian, to leading scientific journals including Cognitive Science. In addition to his award winning writing, Brian is also a Visiting Scholar at the University of California, Berkeley. His most recent book, The Alignment Problem, argues that “when the systems we attempt to teach will not, in the end, do what we want or what we expect, ethical and potentially existential risks emerge. Researchers call this the alignment problem.”
In his talk, Christian noted that, as machine learning expands into an ever-increasing number of areas of our lives, it is “mediating our experience with the world in ways large and small.” He began by identifying several areas of particular concern to researchers and engineers working in machine learning, including limitations in the datasets from which models have historically been developed. Christian offered one particularly vivid example from the field of facial recognition: one of the most popular datasets of faces researchers historically utilized in developing their models was scraped from newspaper articles that appeared in the late 2000s, causing a vast underrepresentation of a number of groups, most notably women of color — a blindspot that ultimately led to disturbing consequences.
Christian identified areas within the field of artificial intelligence where there are reasons for both optimism and hesitancy. Significant successes in the development of artificial intelligence in recent years have been achieved through the increased use of “reward modeling” and “reinforcement learning,” allowing humans to interface with models to identify desirable outcomes, even where those outcomes cannot be easily articulated in mathematical terms. By allowing models to rely on human subjective assessments of what is “better,” models can quickly hone-in on desired outcomes. While this approach has resulted in models that can quickly achieve desired outcomes, the models also necessarily reflect the subjective human viewpoints of those individuals on which the models rely for input. Christian noted that this raises significant questions of equity - when models are being developed around the subjective, often-unarticulated viewpoints of individuals, it is essential that a diversity of groups and perspectives have a “seat at the table” in order to develop models that reflect a diversity of experiences — and therefore limit the myopia that can come from utilizing an underinclusive range of human collaborators.
Among Christian’s most fascinating observations concerned the ways in which artificial-intelligence learning processes based on incentive structures have been shown to mirror the internal learning processes that have evolved in the human mind over millennia. To Christian, this suggests that researchers in artificial intelligence have hit upon something more fundamental than a set of “useful hacks” capable of producing desirable outcomes, but instead are tapping into fundamental truths about the nature of learning. While Christian’s talk was not primarily focused on the future capacities of artificial intelligence (an examination of the present state of the science raises enough ethical dilemmas) one is left wondering, given the enormous progress that has been made in the last few years, what transformative developments (and attendant moral dilemmas) lie just over the horizon.
Tags: Center News
.jpg)