Arjun Subramonian (United States, 24 years old) grew up by chance in California, a place known for glamor of Hollywood and above all for being the mecca of technological innovation. He Golden State It is the birthplace of hundreds of computer industries brought together in Silicon Valley, an environment that led him to imagine in his childhood how computers would power cutting-edge automobiles that humans would drive in the future. But something changed in Subramonian’s perception when he was studying Computer Science at the University of California. At that time he declared himself openly gay and thought daily about the marginalization of the LGBTIQ+ group, which figures 20% less in the technological space. “There are large language models trained with information from the Internet. And we all know that if you are a woman, black or homosexual, that place will be very toxic for you,” she reflects on algorithmic or machine learninga series of unwanted biases that affect generative artificial intelligence (AI) systems.
Machine learning is not free from the biased results that cause hundreds of distortions to fall like darts against groups traditionally excluded from society, which at the same time are underrepresented in the disciplines. STEM, the English acronym for science, technology, engineering and mathematics. Subramonian is one of the main promoters of Queer in AI, an organization that emerged in mid-2020 to make diversity in scientific production visible with activists distributed in more than 47 countries around the world. “Censorship is one of the greatest harms I see in the short term,” he warns.
Subramonian recalls that during the pandemic, activists met on a messaging platform called Rocket Chat. When they were trying to talk to each other, they realized that the term queerfor some reason, was not sending. “We were forced to avoid that word because it prevented messages from being sent,” he says. This occurs, according to Subramonian, because there are AI blocklists that moderate user-generated content so that it complies with a platform’s guidelines and policies. It is a digital process that can filter and manage data from texts, images, videos and even live streams.
“Certain words are often identified as inappropriate,” he explains. Until 2012, for example, Google considered the term bisexual a type of pornography. Data is the raw material of AI and when trained with partial information, the machine can deliver quite limited answers. A situation of this type happened to Ilia S. (Russia, 29 years old), a transgender person and human rights activist based in Thailand who left her native country after the war broke out in Ukraine in 2022. “My voice is very androgynous, the tone often fluctuates and they confuse me with an old woman and in some cases, with my mother,” says the lawyer, who chooses not to reveal his full name to protect his safety.
The most distressing moment he has experienced with an AI system was with his bank, when he lived in Russia. I had to verify a transaction that did not go through. The bank executive began asking him specific and unusual questions such as his place of birth and how much balance he had available in the account. “It’s not you,” was the message he got, since that was the conclusion of the AI voice recognition system. The lawyer then received a threat to block his card for fraud. “I was almost crying because I couldn’t prove it was me,” he says. “Now there is the option to make video calls to compare your face with the photo, but on many platforms this is not the case,” he clarifies.
The censorship of queer
Content biases can cause harmful and misleading results, whether intended or not, on the Internet. Arjun Subramonian had such an experience with his biography. He once came across a text created with generative AI that collected all his information online. The problem was that he had the wrong pronouns, since he is a non-binary person. The error is known as misgendera phenomenon that affects the transgender and intersex population. “The content is monitored and is outside your control. That generates a lot of censorship with the appearance queer of your identity,” he reflects.
For Ilia S., this is a universal problem suffered by trans people who suffer from difficulties of this type. The Russian jurist has been working as an activist since 2013, the same period in which Vladimir Putin promulgated the law against LGBTIQ+ propaganda. Since the measure has been in force, the amount of content on these topics on Russian websites has been disappearing, something that has also affected the book search systems of public libraries. “If you enter certain terms related to LGBTIQ+ content, you can’t find anything because the titles were removed,” he says.
Other flaws detected by experts are in facial recognition systems, which are tools capable of assimilating people’s faces through gender, age and facial characteristics. Ilia’s face is also not read correctly with some of these apps, such as MidJourney. It is a software that generates images from text and now with the function Character Referenceyou can also do it with photos of human faces. In an older version of this program, the face could not be processed.
“Sometimes these apps read my face like an old person or generate monsters. When Midjourney introduced the character reference, it improved a lot,” he says.
Neither good nor bad
Although AI is capable of reproducing homophobia, machismo or racism – a study by the University of Deusto already warned of this in 2023 -, in reality it is neither good nor bad, says Cristina Aranda (Madrid, Spain, 48 years old) , PhD in Theoretical Linguistics and co-founder of Big Onion. “It’s like my washing machine, I decide what program to put on and most of the companies that make decisions are made up of middle-aged heterosexual men, educated at the same universities and who probably have barbecues together,” he ironizes.
The expert, who works with companies to automate processes with language, believes that the technology sector also needs humanities profiles. “There is nothing more philological than the word,” he says. Along these lines, Adio Dinika (Zimbabwe, 33 years old), political scientist and researcher at the Distributed Artificial Intelligence Research Institute (DAIR), thinks that “it is not a tool that invents things”, but rather a product in which a white man’s vision prevails: “Silicon Valley came to determine the AI agenda.”
A few years ago, Aranda took a course on how to help brands position themselves in search engines when he was unemployed and there he had an epiphany. “I had not realized that when we search on the Internet we write words. Google is still a somewhat lexical-semantic search engine,” he says. Aranda refers, in simple terms, to the vocabulary, the hierarchy of meanings and the interpretation of words carried out by the platforms.
80% of the data on the Internet is unstructured, such as social media posts, video surveillance or meteorological data, and it is important that other professionals, such as psychologists or anthropologists, also get involved in the classification process. “These systems are smart, but amazingly dumb because they lack this social intelligence. Come on! They lack streets,” Aranda emphasizes.
There is a technological feudalism and it is the large companies that “are deciding what to do with that artificial intelligence,” he adds. Ilia, for his part, believes that establishing a universal regulatory framework “is very difficult” to avoid these problems. On the contrary, the other experts consulted agree that certain parameters should be established to address these biases.
“Much of the technology we see is developed by corporations based on surveillance capitalism. The population is being subjected to the decisions of these systems and for that reason alone I think it is very important,” concludes Arjun Subramonian.
