The Myth of Neutral Artificial Intelligence

Jasper Zhong

Artificial intelligence systems increasingly affect decisions with tangible consequences for human life : employment, lending, bail, medical outcomes, and sentences to jail. Their common characteristic is being presented in a way that these systems are objective and efficient when it comes to making decisions undeterred by the intricacies of human thought. However, in being applied in operational environments, a whole array of AI system networks ends up with decisions hitting marginalized populations in an economical and social manner. This is neither random nor an isolated event. The impact of these decisions can be expected and replicated with predictable accuracy based on AI systems design and implementation.AI bias defines systematic and directional error, which targets specific groups in a negative way. In this case, systematic error is different because when referring to bias in AI, a bias can be considered a signal for structure based on systematic error. In other words, systematic error can be observed when AI bias affects different dimensions such as those of gender, disability, race, and other social status considerations. AI bias is not a technicality because AI bias results from human decisions embedded in computers.

One of the simplest kinds of bias in AI relates to the nature of the training data. Machine learning algorithms are based on identifying trends among sets of historical data. When this historical data is incomplete, nonrepresentative, or tainted with discrimination, algorithms have no other option but to fit those same trends. The lack of representation is one such simple case. When a particular demographic is underrepresented in small quantities in these datasets, algorithms will train well on them but badly. Such can be observed in facial recognition technology, where systems that were primarily trained using images of white males were able to register accuracy levels of nearly 99 percent in this demographic, but accuracy in black females dropped to just 65 percent, while in another large study, error rates were 34.7 percent in dark-skinned females compared with 0.8 percent in light-skinned males. Such imbalances did not arise out of technological flaws, but rather were a function of a biased dataset composition. Moreover, selection bias will perpetuate this problem. The training set will all too frequently represent a historically valued population. A selection algorithm using resumes in a male-dominated population will hence associate achievement with qualities coded in males. Credit rating algorithms, in modeling higher-income populations, do not generalize to lower-income populations. In such algorithms, a population is not incorporated to which it will serve.