From Fair, Accountable, Transparent Technology to a Science of Ethics
by Paul Hope
Data, algorithms, machine learning, and artificial intelligence are at work in all the digital platforms we use -- including this very Substack. In 2018, partly in response to a series of scandals around racial bias in recidivism prediction and housing advertising algorithms, the first Association of Computing Machinery conference on Fairness, Accountability, and Transparency took place in New York. This annual conference was bigger than ever this year in Montreal.
The conference attracts an eclectic mix of university and industry AI researchers and social scientists, policy wonks from government and think tanks, and civil society activists. This year, there were four keynote lectures, hundreds of paper talks, and close to 50 tutorials, spread over four days. Topics ranged from addiction to digital services, consumer privacy, regulatory mechanisms for AI systems, and fairness interventions to improve outcomes of digital systems for marginalized groups. One steady theme of the conference has been a long-known issue: that systems designed for the “normal” majority of people tend to impose hidden costs on those that are more marginal, and this contributes invisibly to structural inequality over time.
When this conference began less than ten years ago, there were not many places where those with sincere concern for the impact of technology on society could convene, share research, and form actionable plans to align the digital systems we depend on with principles of justice. Prior to conferences like it, technical expertise was typically seen as value-neutral and was separated from humanistic concerns. But as digital systems are increasingly intertwined with every aspect of life, this distinction is more and more untenable. A new interdisciplinary field of research arose out of practical necessity.
As we work towards a practical and scalable understanding of designing artificial intelligence that is aligned with social values like fairness, accountability, and transparency, we learn something quite profound. Researchers in this field are, out of practical need, developing a science of justice and ethics that can be encoded into software systems or public policy that regulated technology services. An example of this is the rapid development of a deep formal theory of “fairness.” There are, somewhat famously, at least 21 different mathematical definitions of “fairness,” and, correspondingly, a wide range of ways in which a given social setting can be “unfair.” The most mature theories of fairness demand a careful accounting of cause and effect -- a clear representation of what factors influenced which decisions, and why. These causal pathways to fairness are the results of how organizations, often without realizing it, make decisions, or delegate them to digital tools.
Ultimately, the ethics of digital technology opens up in an ethics of institutional design. How organizations convene, include or exclude people in decision-making, enact bureaucratic procedures, and decide to use technologies has implications for ethics and justice. We can hope that, in the future, these principles will be broadly applied, and our more careful self-organization can improve the world.