Can AI be sexist?

AI has been shown to have the capability to reproduce human errors.
The human brain is susceptible to a bias called heuristics, one example is the representative heuristic. This heuristic contributes to probability errors that are common in humans, especially when fed limited information. A famous example by Tversky and Kahneman is the Linda Case. Participants are given a description of a woman and are then asked which of two options are more likely.

They are told, “Linda is 31 years old, single, outspoken, and very bright. She majored in philosophy. As a student, she was deeply concerned with issues of discrimination and social justice, and also participated in anti-nuclear demonstrations. Which is more probable? (1) Linda is a bank teller; (2) Linda is a bank teller and active in the feminist movement.”

Similarly, researchers from Carnegie Mellon University and the International Computer Science Institute created a program called AdFisher to see how Google targets job seekers depending on their gender. They found that users listed as men were shown high paying executive jobs 1,852 times but only 318 times to women, almost fives times less than men.

Why is it that such a phenomenon could be reproduced in an entirely man-made mechanism? For that plain reason, entirely man-made. The lack of a woman's perspective in the development of AI is a huge issue that could lead to unfavorable repercussions should the female perspective be ignored. This is not to say it is the male programmer’s fault for being male, but the fault lays at the foot of social construct which creates a barrier for women who strive to pursue STEM jobs.

Due to the complexity of the human brain, it is incredibility difficult and demanding to simulate a range of human processes. However, even our brain uses shortcuts. Heuristics – which we mentioned earlier – are only one of the shortcuts we use. Programmers use these shortcuts to replicate human recognition, and what better shortcut than stereotyping, right? Before you get too confused, stereotyping is widely misconceived to become affiliated with issues of sexism, racism, etc. However, this is a very complex mechanism, and AI has not been doing great with it so far. A current example is when Google Photos mistakenly classified black people as gorillas. This incredibly discriminatory problem has at least one up-side to it, it resurfaced the issues behind underrepresentation in programming. AI learns through data, algorithms, and experience, however, how will it learn to discriminate a wide range of faces if they are not considered during the programming stage? 

Can AI be biased?

AI can’t really be biased, per say, since they cannot “want” to believe something, they just, sort of, do. To even say it can be programmed to believe would be difficult to confirm since belief suggests some kind of agency in thinking. However, programmers can reproduce certain values through the machine, but those values would not necessarily be from that of the robot, but that of the programming. Until we create a fully self-aware and learning AI, that is. 

How is this different than the sexism or racism argument, you might be wondering. Well, let’s assume that the programmers tried their hardest to keep the AI unbiased, but it continues to represent unbalanced values. The programmers may have inadvertently favoured their bias and programmed it in, whereas with the subject of gender and race, this issue is what is not there rather than what is suppressed.

Biases are influenced by concepts such as fear, anger, passion, and other emotions aside from logic. Computation can only take us so far in this respect, how would one quantify humiliation, for example. Or one step further, the concept of schadenfreude, which is defined as “the experience of pleasure, joy, or self-satisfaction that comes from learning of or witnessing the troubles, failures, or humiliation of another”. Additionally, AI has difficulty weighing something such as importance. AI recognizes all mistakes as equal, but of course that’s not true. For example, asking Siri to send a text saying “get here fast” is much more problematic when sent to 911 rather than when it’s sent to the wrong person on your contact list. If AI cannot differentiate the significance of the two mistakes, it will have difficulty learning value judgement. In this sense, it could be said that it is unbiased, since it treats the situations similarly, but perhaps not in such a beneficial way.