Generative AI has a problem with women. Particularly, with the way AI systems they create images that represent (or should represent) beauty female.
The last example, in order of time, is that of Smash or Pass, an online platform that allows, along the lines of Facemash, a sort of ancestor of Facebook, to evaluate female images positively or negatively. Photos of people who don’t exist, generated by artificial intelligence. By casting the vote, the AI should be able to understand the user’s tastes and to create increasingly pleasant images for him (or her).
Aside from the architecture of the platform itself, which is sexist right from the design, what is striking are the images. Enormous breasts, faces without imperfections, very young womenalmost all white: generative artificial intelligence, as already demonstrated by other experiments the da some generations of the Lensa apphas a decidedly stereotyped idea of female beauty.
Artificial intelligence AI between racism and discrimination: two years later, we are worse off than before by Emanuele Capone 27 March 2023
Beauty in the time of artificial intelligence
“The issue starts from afar – according to what he explained to us Giada Pistilli, Ethics Manager of Hugging Face, a US company that develops tools for machine learning – I am reminded of the case of the photo in Playboy of Lena Forsén, a Swedish model, used since the 1970s as a guinea pig for digital image processing. It is an example that helps to understand a direction: in the world of artificial intelligence and information technology, a certain standard of beauty has always been presentright from the origins.”
To explain why Lena’s image had become an industry standard, David C. Munson, direttore della rivista IEEE Transactions on Image Processing, explained that while it was a good test image due to its detail, shading and texture, the image depicted an attractive woman. A circumstance much appreciated by the image processing research community (made up mostly of heterosexual males).
In short: on the one hand, women are suffering from years of underrepresentation in the tech world: according to a 2023 report, for every female professional, there are 4 male ones; on the other hand there is the very functioning of generative artificial intelligence systems. Who, as is known, are trained to create new media starting from an enormous amount of examples, from databases through which they learn to know the world and to link the words provided in input, the famous prompt, to the output.
“Using words to describe an image is not a process that arises with prompts – he recalled Antonella Sbrilli, art historian at Sapienza University – It is a literary tradition, which involved the use of pictorial/sculptural models in the text, so that readers could imagine features and attitudes; there are many examples: hair, face shape, neck curvature, colors are recalled through comparison with Botticelli, Velazquez and so on. Looking at them now, one could say that these descriptions (as beautiful as the Venus de Miloaustere like a fresco by Giotto and the like, ed.) are like prompts, which draw from the repertoire of shared visual culture to generate mental images”.
However, the question concerns the way words are linked to the images generated or, to put it another way, the visual culture available to a society. What does beautiful mean, when the representations I have available identify as such almost exclusively women who respond to very specific stereotypes?
“The concept of beauty has always been represented, since ancient times, in both the female and male bodies – he reminded us Marinella Belluatimedia sociologist and one of the promoters of public engagement project of the University of Turin AI Aware – What has happened is that with mass society, the media and pornography there has been an orientation of the definition of beautiful, pleasant, desirable, especially as regards women. In particular, long before artificial intelligence, a canon was defined, that of curvy, languid and sexualized woman”.
Axel Wahlström, who created Copy, the first magazine with images entirely generated by AI, has revealed to Business Insider that he had difficulty creating photos of ordinary women, that they weren’t too perfect: “I couldn’t understand why. Whatever the prompt, the result was always very similar: perfect and stereotyped models – he explained to the American magazine – I believe that our experiment was also a way to demonstrate the image of beauty that we had and spread for the last 40 years”.
“This canon it also settled in the unconscious of artificial intelligence, trained to think of those bodies as beautiful, on the basis of the material provided as input by the programmers – added Belluati – The issue is that it is a stereotypical, masculine and chauvinist way of imagining the female body. The risk is that it will strengthen model that is increasingly difficult for anyone to achieveand that non-compliant bodies remain marginalized.”
Artificial intelligence OpenAI will host its first developer conference in San Francisco by Gabriella Rocco 07 September 2023
The solution, or the birth of a new imagination
It is not easy to find a solution. The AI community is divided on the topic: some even talk about eliminate women from demos and early product releasesto avoid the risk of stereotyped representations and there are those who try to imagine a different solution, which starts from system training and transparency.
“We must not forget that technologies are linked to the society we live in: if the society is sexist, it is difficult for the AI, trained on the images that this same society produces, not to be so – is Pistilli’s reasoning – Surely an effort can be made to use databases that include more realistic and start a more open dialogue on the topic.”
Because it is not said that artificial intelligence necessarily has to be reproduction of the world as we know it: “The complexity and density of these phenomena suggests that they cannot be assessed clearly – is Sbrilli’s suggestion – The metamorphoses in the production and reuse of images can simultaneously strengthen and exaggerate, but also empty and replace canons, beauties , prejudices and imaginations”.