The same thing happened with the first crash test dummies – dolls simulating the human body used to measure the impact of accidents. They date back to 1949, but we’d have to wait until 1996 to have the first “pregnant” one. In 2020 there was still no model recreating the average female’s body size and anatomy, making women who wear seat belts 73% more likely to be seriously injured in a head-on crash than men, according to a University of Virginia study. We can see, therefore, that these imbalances have a social cost – all because women are not taken into account in these innovations.

EMOJIS WITH A MALE FACE

Another example of the “male standard” is that of emojis, those cute little ideograms that have crept into our daily lives through mobile phones and social media. When these characters of Japanese origin were incorporated into the Unicode Standard, gender was not established. The Unicode Consortium set an alphanumeric code at the global level, and each technological platform represented the code however it chose. What happened? Most made them to resemble men. For example, the Unicode Consortium defined that the code U+1F4BB corresponded to a person working on their laptop, and both Android and iOS (to give two examples) showed a boy with a computer. Thus, in 2016 they introduced gender to be able to indicate whether the person on their computer was a woman (code U+1F469+U+1F4BB) or a man (U+1F468+U+1F4BB).

BIASES IN ARTIFICIAL INTELLIGENCE

As we can see, these gender biases cross over into the digital realm, just as they exist in the analogue one. In fact, in the age of Artificial Intelligence, Machine Learning algorithms not only fail to prevent human error stemming from biases, but they can even worsen discrimination, and are reinforcing many stereotypes. Techniques like Machine Learning process large volumes of information to detect patterns, formulating rules based on them. If these huge amounts of data are generated by a biased society, technology only magnifies those biases. An example of this can be seen in Google Translate. This system learns from the translations that are done from inside its own tool, or by processing digital publications. If we carry out translations from Turkish, a language that has no grammatical gender, we can see how inequalities are replicated. In this language, there is no “he” or “she.” Thus, if you translate Turkish into English, it is the technology that must decide the gender to assign based on what it has previously learned. This is why we find gems like “she’s a cook,” “he’s an engineer,” “she’s a nurse,” “he’s a doctor,” “she’s lazy” and “he’s a hard worker.”

Something similar happened with a predictive modeling algorithm that was not able to identify men cooking in photographs. When seeking to understand why, the algorithm’s creators realized that in the main image banks from which machines learn, 77% of the photos in which people were shown cooking featured women. Thus, the technology inferred that if someone appears in a photo with a frying pan in their hand, it must be a woman.

We have another similar case that got its start in 2003, when the names of the people, places, and organizations that appeared in online news began to be collected manually. The idea was to build a database with which algorithms could generate models. The CoNLL-2003 repository was created with this material, made available to the scientific and research community, and used in countless applications (it has been cited in more than 2,500 papers). The problem is that in 2020 it was discovered that there are five times more men’s names than women’s, and the systems “trained” for 17 years with this data fail more when identifying new women’s names.

INTERSECTIONALITY

And, if this were not enough, upon adding other key factors to the gender component (such as race, age, ethnicity, class, disability, etc.), we have the perfect intersectionality cocktail, revealing how the various social categories bring about very different forms of oppression, resistance, and privilege when they intersect. Intersectionality, then, allows us to notice elements that we miss at first sight, and shows us that if a person is immersed in a large number of oppressed identities, those will end up causing multiple oppressions as well. We can see an example of this with the API that Google has to combat “toxic behavior” on the Internet: Google’s Sentiment Analysis, used by many chatbots and webpages. It is a tool that, for any given text, returns a “toxicity score” or, in other words, “something with negative sentiment,” to make it easier for a platform to automatically discard comments with a high score (for example, an online newspaper). Well, here are the alarming results that the tool yielded in 2017: if the comment included the phrase “I’m a man,” the toxicity level was 20%; however, if the comment included “I’m a woman,” it went up to 41%. If it found the text “I’m a black woman,” the score would rise to 85%; and with “I’m a black lesbian woman,” it would shoot up to 87%. The reason is that most of the comments with insults that had been collected manually included some combination of those concepts.

Therefore, we see that we are facing an obvious social challenge: looking at the technology that is shaping our society today from the perspective of diversity (gender, age, race, etc.) so as not to establish a single model as the standard on which everything is built and that is not representative of everyone. The challenge is great, as the data indicates that there is much to be done in areas such as Engineering. If we look at the evolution of female students enrolled in Computer Science in Spain from 1985 to 2019, we see that in the first year for which data is available (the 1985-86 academic year), 30.12% of those enrolled were women. By 2019-20, that figure had plummeted to 12.74%. Sheryl Sandberg, the former Chief Operating Officer at Facebook, once said, “when you write one line of code, you can affect a lot of people.” This is why half of humanity cannot be left out of the technological designs that shape our society, both today and tomorrow. That society cannot be looked at from the perspective of a single standard (the male one).