In 1993, biologist Anne Fausto-Sterling proposed that our two-sex system should be replaced by five: males, females, herm (“real” hermaphrodites), serm (male “pseudohermaphrodites”) and serf (female “pseudohermaphrodites”). Beyond the fact that this idea had a certain ironic tone, what I wanted to highlight is that the labels “man” or “woman” not only have a scientific component, but also a sociological one .
Therefore, intersex covers several conditions that give rise to bodies with a mixture of male and female parts. Currently, various types of intersex have already been described depending on their physiological origin, such as androgen insensitivity syndrome or congenital adrenocortical hyperplasia .
Thus, given that determining a person with a label such as “man” or “woman” is already difficult for human intelligence, it is even more difficult for artificial intelligence (AI), whose facial recognition algorithms are victims of sexual biases.
Cloud vision
Google Cloud Vision API is a tool in the Google cloud that allows you to analyze a large number of images and extract valuable information from them to understand their content. The Face Detection feature, for example, aims to detect human faces in images.
However, since a person’s gender cannot be inferred from appearance alone, in order to avoid sexual biases and other biases, Google Cloud Vision will stop tagging photos of people as “male” or “female” and will opt for the generic “people”.
It is not the first time that Google has faced the problem of correcting sexual, racial or other biases in its various software tools. Biases in artificial intelligence are a widely debated issue among developers, since the data used to train algorithms, whether faulty, incomplete or simply also victims of bias, can finally reinforce unfair situations and prejudices. One example is facial recognition algorithms that misidentify black people more often than white people .
In 2015, for example, a software engineer pointed out that Google Photos’ image recognition algorithms were categorizing his African-American friends as “gorillas.” Since the engineers and programmers have not yet known how to refine the algorithm, the option to identify “gorillas” in the images was simply blocked.
Algorithmic bias , then, occurs when a computer system reflects the values and biases of the humans who are involved in coding and collecting the data used to train the algorithm. However, some developers do not attach much importance to this kind of bias and consider that these concerns respond more to a fashion associated with political correctness than to a purely technical aspect. According to his approach, if a system is able to correctly identify a person most of the time by their sex or ethnicity, just as a human being does, then we must be able to use that information even if there is a small quota of mistakes .
