The lack of transparency in algorithms and the data they use is a matter of current interest. In addition to a nontransparent process, results are also nontransparent, lacking explanation or justification. These problems lead to unfairness and bias, as illustrated by Amazon’s algorithmic resume screener that discriminated against women. Federal action has stalled though there are regulatory guidelines on fairness and nondiscrimination in artificial intelligence (AI) application outcomes.
Corporations can address these matters by implementing best practices in their algorithmic application areas. But Amazon has shown best practices may embody unfairness and bias. Relying on diversity in development and metrics can help expose and eliminate these problems, as well as improve inclusiveness by widening the potential market. While some corporations are leading, others are lagging, particularly those with large, embedded, and long-standing systems that may incorporate algorithmic and information bias. Dealing with corporate lag is another significant problem.
Ultimately, bias is a human problem independent of mechanisms realizing bias. Dealing with bias requires clear-eyed self-reflection by being observant and thoughtful about relevant histories. In this, algorithms will always be a lagging reaction to human activity.
This short article lays out these matters and the attempts at dealing with them. Although philosophical considerations appear in the first paragraph, the article is practical, which is useful but unfortunate. The thinking around algorithmic transparency is fuzzy and unclear, and it is hard to see how large-scale solutions are possible without clarity and agreement on what the problems are.