It’s little wonder, as Kalinda Ukanwa asserts in “Algorithmic bias isn’t just unfair — it’s bad for business,” that artificial intelligence “simply [replicates] our existing prejudices.” After all, when it comes to bias, the weakest link in AI development is the human coder of those algorithms.
The designer carries along to the task his or her own baggage of predispositions that skew decision-making, often for the worst. No matter the care to make the algorithms bias-free, at least some of the designer’s prejudices insinuate their way into the product.
Biases’ roots run deep, but not all are mere spawn of the biases harbored by the developer. As Ukanwa details, in some cases, as part of machine learning, prejudicial decision-making stems from the AI recognizing patterns in data related to an institution’s past decisions and deciding that’s the model to replicate.
Of Ukanwa’s three remedies to the use of AI in decision-making, business, and other contexts bearing heavily on people’s well-being, I suggest that the first — to “continually audit” algorithms and expunge quirks — is the most beneficial.
I might also urge that, proactively during design, developers resort to multidisciplinary, multiethnic, multiracial, multigender “murder boards” of critical reviewers to diagnostically examine algorithms from diverse social perspectives before product launch.
Keith Tidman
Bethesda, Md.