Is AI Sexist?
In the not-so-distant future, artificial intelligence will be smarter than humans. But as the technology develops, absorbing cultural norms from its creators and the internet, it will also be more racist, sexist, and unfriendly to women.
Heather Roff, an artificial intelligence and global security researcher at Arizona State University, cannot shake her trepidation about the future. Her office shelves are replete with titles like Rise of the Robots, Wired for War, and Moral Machines. Alongside those books, the research scientist with the school’s Global Security Initiative also keeps copies of War and Gender, Feminism Confronts Technology, and Gendering Global Conflict. On one shelf, a magnet reads, “Well behaved women rarely make history.” A vintage poster hangs on a nearby wall with a half-naked, barefoot woman riding a missile and the words “Eve of Destruction.”
Also a senior research fellow at the Department of Politics & International Relations at the University of Oxford, Roff recently began working under a grant for developing “moral AI.” She is concerned with how representations of gender are becoming embedded in technology and expressed through it. Gender, race, variations in human behavior — none of this is easily encoded or interpreted in artificial intelligence. In a machine, a lack of diversity manifests through an interpretation of a set of codes. “It’s like a data vacuum sucking it all in, looking for a pattern, spitting out a replication of the pattern,” Roff says. It cannot distinguish whether conclusions from learned patterns violate moral principles.
A pattern can begin on the simplest scale, like an internet search. Recently, researchers from Universidade Federal de Minas Gerais in Brazil examined algorithmic notions of desirableness of women on Google and Bing in 59 countries around the world. They queried the search engines for “beautiful and ugly” women, collecting images and identifying stereotypes for female physical attractiveness in web images. In most of the countries surveyed, black, Asian, and older women were more often associated through algorithms and stock photos with images of unattractiveness, while photos of young white women appeared more frequently as examples of beauty.
The researchers suggest that online categorizations reflect prejudices from the real world while perpetuating discrimination within it. With more people relying on burgeoning amounts of information available through search engines, designers turn to algorithms to sort out who sees what. When those algorithms are not transparent to the public, why and how a system settled on selecting a particular image or advertisement can remain a mystery. Ultimately, this reinforcement of bias between the internet and its users can exaggerate stereotypes and affect how people perceive the world and their roles in it.
InferLink Corp. of El Segundo, California, draws on data, artificial intelligence, and machine learning for the government, universities, cybersecurity firms, and other companies. It analyzes behavior on social media, layering its data with algorithms infused by websites and psychological and linguistic studies. “We can take Twitter, Reddit, and blog posts and turn them into a set of demographics and interests,” chief scientist Matthew Michelson told his audience at the Las Vegas conference during his talk on “Discovering Expert Communities Online Using PSI 14.”
Like other similar programs, InferLink algorithms incorporate research into how men and women express themselves and speak differently online. “Men use more declarative verbs,” Michelson told me after his talk. “Women are more descriptive.”
Research over the last three years has uncovered gender differences in social media language. A recent study in PLOS One, the open-sourced, peer-reviewed journal of science and medicine, reviewed 67,000 Facebook users and found that women used “warmer, more compassionate, polite” language in comparison with men’s “colder, more hostile, and impersonal” communication.
Female users, the study notes, more often use words associated with emotions like “love,” “miss,” and “thank you,” and emoticons of smiles, frowns, and tears. Meanwhile, male users are more inclined to swear; talk about management, video games, and sports; and include more references to death and violence. Previous studies of spoken and written language have showed that women tend to hedge more, using words like “seems” or “maybe.”
But once you get into making inferences about gender, race, or socioeconomics based on any of these algorithms — whether for things like marketing or policy advising — Michelson says using the technology gets into touchy territory. This is how women might be targeted unequally for financial loans, medical services, hiring, political campaigns, and from companies selling products that reinforce gender clichés. “We don’t want to unleash something we can’t undo.”
This is also how nontransparent algorithms can become indirect tools of discrimination and directly affect women’s livelihoods. There are already algorithms that are more likely to show online advertisements for high-paying jobs to men. Google image searches for “working women” turn up lower rates of female executives and higher rates of women in telemarketing, contrasted with women who actually hold such jobs.
Roff warns that women could lose out on opportunities because of a decision that an algorithm made on behalf of them, one “that we cannot interrogate, object to, or resist.” These decisions can range from “which schools children go to, what jobs we can get (or get interviews for), what colleges we can attend, whether we qualify for mortgages, to decisions about criminal justice.”
Algorithms could one day target women personally, Roff explains, telling them what is normal. “[They] will manipulate my beliefs about what I should pursue, what I should leave alone, whether I should want kids, get married, find a job, or merely buy that handbag,” she says. “It could be very dangerous.”