Who Is More Sexist: Humans or Artificial Intelligence?

Anne Fehres and Luke Conroy & AI4Media / Better Images of AI / Data is a Mirror of Us / CC-BY 4.0

The ongoing debate over whether artificial intelligence (AI) tools are sexist has grown even stronger recently with AI-generated images that seem to be everywhere. For example, images produced by many different AI tools such as DALL-E, Midjourney, Bing Image, Leonardo AI, and even Canva have brought sexist elements such as stereotyping and objectification to the agenda.

To discuss what kinds of sexist elements an image may contain, we spoke with Ceren Suntekin and Ekin Taneri, who work in the field of visual design, and also evaluated how these sexist elements are reproduced in AI tools with Sabire Sanem Yılmaz, who specializes in artificial intelligence and ethics. While Suntekin and Taneri pointed out that sexist methods are highly technical but originate from a patriarchal society, Yılmaz said that the sexism in images generated by AI tools stems from image stock datasets.

Colours, Fonts, Poses, Representation: Elements of Sexual Abuse

Visual design methods may fail to include diversity. In the resource Extra Bold, Kaleena Sales says that the dominant movements and rules in graphic design consist of Western-centered and hierarchical movements such as Bauhaus, Constructivism, and the International Typographic Style, which leave no room for difference.

Ceren Suntekin

Suntekin, who specializes in children’s rights and says she found a means of self-expression through visual design, said that visual designs are products of the human mind and can become sexist through the fonts, language, colours, icons, shapes, placements, emphasis and images used by people. She gave the example of posters on sexual abuse:

“Generally, the images used in such posters show a young girl whose mouth is covered by a man who makes us think he is a construction worker or a migrant. The girl is terrified. Dark colours, such as black and red, are used in the poster. The announcement text is written in fonts that look as if they belong in tabloid news.

This poster presents only women or girls as survivors of sexual abuse. It presents only socioeconomically disadvantaged men as perpetrators of sexual abuse. Although abuse is generally committed by someone closest to the child, it creates the image that it comes from a ‘stranger’ the child has never met. These dark posters create the perception that sexual abuse ends a person’s life. They create the perception that sexual abuse can only be physical and involve bodily contact. Consequently, such posters reinforce myths that already exist in society, distort reality, and create discriminatory language through gender roles.”

Clarote & AI4Media / Better Images of AI / AI Mural / CC-BY 4.0

Taneri, who works as a graphic designer at an agency, agrees that many details in visual design, from typography to colour selection, make images sexist, and says that women’s representation in the world of design remains limited to thin fonts and the colour pink.

“Thin fonts, bold fonts, and the hierarchy within a design generally determine sexism. Sometimes the visual itself is designed to resemble a woman’s body, and sexism is created in this way. The target audience is generally men. There is a great deal of sexist visual production in advertisements. It is especially common in beer advertisements.”

Could Fast Consumption Be the Cause of Discrimination?

Taneri said that images produced by social media professionals and advertisers are designed to be consumed quickly and are perceived within as little as three seconds, adding that agencies continue to use whichever visual attracts people’s attention most quickly. She said this has made graphic design increasingly uniform.

Suntekin explained how this standardization takes place with an example and described how it can be challenged:

“When illustrating a disabled person, I sometimes have to depict someone using a wheelchair. In fact, not every disability has to be visible from the outside, but in order to attract attention we may resort to standardized representations that carry the same meaning for everyone. I try to move beyond this by diversifying these representations.”

So, what do these sexist and discriminatory elements in visual images have to do with images generated by artificial intelligence?

Sabire Sanem Yılmaz said that the sexism we have seen in visual culture until today continues in AI-generated images in the same way it has been reproduced in the past. For example, she observed that DALL-E continues to assign colours to genders.

Sabire Sanem Yılmaz

AI Produces Female Housekeepers and Male Engineers

A study conducted in Spain, Gender Stereotyping in AI-Generated Images, found that DALL-E 2 generates images that stereotype women and men in relation to occupations. The images generated by searching for 37 different occupations in DALL-E 2 assign professions to women and men, in other words, they reproduce stereotypes. The study found that 21.6% of the images assigned certain occupations exclusively to women, while 37.8% reproduced male occupational stereotypes.

When users search for occupations such as housekeeper, primary school teacher, teacher, singer, tailor, hotel manager, and secretary, the generated images depict women. Meanwhile, searches for carpenter, taxi driver, truck driver, airline pilot, mechanic, construction worker, soldier, engineer, barber, police officer, banker, computer specialist, politician, and priest or religious leader generate images of men.

The study conducted on DALL-E 2 shows that stereotypical gender roles are reproduced in images in 59.4% of the 37 occupational categories.

Smoother, Scar-Free Women Without Signs of Lived Experience

So why do these tools reproduce stereotypes? Why don’t they generate more equal representations? Shouldn’t technology be more objective?

In response to these questions, Yılmaz pointed to the datasets used in AI tools:

“The source of the data used in these AI models, its quality, and whether it meets transparency criteria are extremely important. The datasets are shaped by what societies understand about the roles of women and men. Data means people — it means their DNA. Female figures are depicted as overly sexualized, with large breasts, but there is something else we should not overlook. Women are transformed into smoother figures, without scars, without signs of lived experience. In fact, this is the most dangerous part. It creates a model of womanhood that is stripped of posture, charisma, natural talent, or the meaningful expressions that settle on a face over time.”

The datasets used by image-generating AI tools also include hundreds of thousands of stock images. So why do companies use these discriminatory image stocks? Isn’t it possible to use more inclusive datasets that do not reproduce discrimination and sexism?

Yılmaz explained this in parallel with state policies:

“I do not believe that AI tools can produce outputs that differ from the policies of the state in which they are developed. It is because of this concern that phenomena such as discrimination, prejudice and manipulation have emerged. In studies carried out by experts to eliminate or mitigate the prejudice and discriminatory content contained in datasets, various quantification and minimization techniques are being tested.

What matters here is that AI developers adapt to and commit themselves to legal and international regulations in response to existing risks. The European AI Act adopts a risk-based approach, and we should not ignore the possibility that high-risk AI systems may be sexist.”

Clarote & AI4Media / Better Images of AI / Labour/Resources / CC-BY 4.0

What Should Be Done?

Suntekin, who produces rights-based visual representations and illustrations, underlined that there is a need to establish a stronger connection between technology and gender equality in response to sexism in AI-generated images. Yılmaz added a more legal perspective, saying that without regulation, these uses could reach a much more dangerous level. She argued that if the datasets used by these tools are overseen by trustworthy bodies, sexism may not be reduced to zero, but more dangerous forms of discrimination can be prevented.

Yılmaz also shared with us for the first time in Turkey a letter of appeal that she helped prepare together with a group of activists, academics, researchers, civil society advocates and experts in the field, addressed to many European institutions, including the European Parliament.

Through this letter, the activists drew attention to issues that require action, including preventing discrimination in Generative AI models, biased AI models against women, inequalities of opportunity resulting from lack of access to technology, disadvantageous decisions affecting women made by AI decision-making systems, and AI systems designed without taking vulnerable individuals into account.

In the letter, the activists made the following call regarding Generative AI models:

“The structure of the datasets used to train Generative Artificial Intelligence algorithms should be reviewed before they spread the biases found in other artificial intelligence tools. We call on the developers and users of these technologies to proactively commit themselves to the principles of fairness, transparency and inclusiveness for all individuals, including vulnerable persons.”

“We call on the European Union to take the lead in developing ethical international standards and guiding principles capable of protecting the rights and freedoms of vulnerable persons, thereby ensuring the protection of all humanity.”

Author: Asmin Ayçe İdil Kaya

Original Source: https://9koy.org/kim-daha-cinsiyetci-tasarimcilar-mi-yapay-zeka-mi.html