Free beta access. 50 free AI tokens each week. Optional top-ups available. See what is included
Opening Word document…

AI reproduces the subjectivity of the people that create it---statistically that is white, male, and wealthy.

Roff, 16 [Heather M. Roff works for the global security imitative at Arizona State University, 1-11-2016, accessed on 6-21-2022, Taylor & Francis, "Gendering a Warbot", https://www.tandfonline.com/doi/abs/10.1080/14616742.2015.1094246]/ISEE

This brings us full circle to the problem of which objects and concepts get represented in the dataset. If the programmers, who are predominantly white, male, upper-middle class and academic, equip an AI with these perspectives, then it is not difficult to see how gendered constructions will reproduce themselves (AAUW 2015). A particular masculinity might be reproduced, or a particular view of femininity may not be modeled at all (Adam 1998, 23). In either case, the AI is clearly not “gender neutral.” Sub-symbolic systems fare no better. In these “bottom–up” systems, the AIs are left to figure things out on their own. The AI is a blank slate, where programmers envision creating an intelligence on par with a child’s brain.20 By emulating the learning processes of a child, from adolescence through adulthood, one is able to create a generally intelligent AI (Storrs Hall 2007, 137). Aside from the ethical considerations about an AI-baby, there are other technological limitations and hazards about bottom–up systems. In particular, because there are no preprogrammed constraints (e.g. do not kill noncombatants) and it is a learning system, it has the potential to learn the “wrong thing” (Wallach and Allen 2009, 110). Programmers acknowledge the potential for such “bad behavior” and have taken seriously the AI-baby approach where a programmer oversees the “education” of the AI. Yet this does not assuage fears of a gendered AI. In fact, it would almost guarantee not only a masculinist AI, but would also pervert the very nature of learning. The AI would have very limited social interaction: it lives in a lab, does not go to school, interacts with one type of person and is limited to whatever interaction its “caregiver” decides to give it (Weber 2014). This is not to say that there are no women in robotics or AI labs, but that the culture permeating computer science and robotics is predominantly masculine, and that the women have consistently remained a mere quarter of the population of computer scientists since 1960 (Burleigh 2015; AAUW 2015). Finally, the statistical approach will also produce gendered AI. It works off of basic probabilistic mathematical models of how the world operates. The statistical approaches will utilize certain kinds of modeling, such as Bayesian inference or networks. In a Bayesian inference, the model will predict with x probability that a statement is true, whereas a Bayesian network is a graph “whose nodes represent statements or measurements,” and the “links between the statements or measurements correspond to the Bayesian-inference steps with the appropriate probabilistic models” (Storrs Hall 2007, 156). Of course much depends on the model, the available data and the scope of the problem to solve. For the machine to solve a problem requires “taking enough statistics from enough evaluations” to first form the model, and thus requires either interacting many times in a particular situation or a large enough data set from which it could draw to be prepared to act with any reliability in a new context.

Sign in to recut, update, or save this card.